Upload 48 files

.flake8

@@ -0,0 +1,3 @@
+[flake8]
+select = E3, E4, F
+per-file-ignores = roop/core.py:E402

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+docs/screenshot.png filter=lfs diff=lfs merge=lfs -text

LICENSE

@@ -0,0 +1,661 @@
+                    GNU AFFERO GENERAL PUBLIC LICENSE
+                       Version 3, 19 November 2007
+
+ Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
+ Everyone is permitted to copy and distribute verbatim copies
+ of this license document, but changing it is not allowed.
+
+                            Preamble
+
+  The GNU Affero General Public License is a free, copyleft license for
+software and other kinds of works, specifically designed to ensure
+cooperation with the community in the case of network server software.
+
+  The licenses for most software and other practical works are designed
+to take away your freedom to share and change the works.  By contrast,
+our General Public Licenses are intended to guarantee your freedom to
+share and change all versions of a program--to make sure it remains free
+software for all its users.
+
+  When we speak of free software, we are referring to freedom, not
+price.  Our General Public Licenses are designed to make sure that you
+have the freedom to distribute copies of free software (and charge for
+them if you wish), that you receive source code or can get it if you
+want it, that you can change the software or use pieces of it in new
+free programs, and that you know you can do these things.
+
+  Developers that use our General Public Licenses protect your rights
+with two steps: (1) assert copyright on the software, and (2) offer
+you this License which gives you legal permission to copy, distribute
+and/or modify the software.
+
+  A secondary benefit of defending all users' freedom is that
+improvements made in alternate versions of the program, if they
+receive widespread use, become available for other developers to
+incorporate.  Many developers of free software are heartened and
+encouraged by the resulting cooperation.  However, in the case of
+software used on network servers, this result may fail to come about.
+The GNU General Public License permits making a modified version and
+letting the public access it on a server without ever releasing its
+source code to the public.
+
+  The GNU Affero General Public License is designed specifically to
+ensure that, in such cases, the modified source code becomes available
+to the community.  It requires the operator of a network server to
+provide the source code of the modified version running there to the
+users of that server.  Therefore, public use of a modified version, on
+a publicly accessible server, gives the public access to the source
+code of the modified version.
+
+  An older license, called the Affero General Public License and
+published by Affero, was designed to accomplish similar goals.  This is
+a different license, not a version of the Affero GPL, but Affero has
+released a new version of the Affero GPL which permits relicensing under
+this license.
+
+  The precise terms and conditions for copying, distribution and
+modification follow.
+
+                       TERMS AND CONDITIONS
+
+  0. Definitions.
+
+  "This License" refers to version 3 of the GNU Affero General Public License.
+
+  "Copyright" also means copyright-like laws that apply to other kinds of
+works, such as semiconductor masks.
+
+  "The Program" refers to any copyrightable work licensed under this
+License.  Each licensee is addressed as "you".  "Licensees" and
+"recipients" may be individuals or organizations.
+
+  To "modify" a work means to copy from or adapt all or part of the work
+in a fashion requiring copyright permission, other than the making of an
+exact copy.  The resulting work is called a "modified version" of the
+earlier work or a work "based on" the earlier work.
+
+  A "covered work" means either the unmodified Program or a work based
+on the Program.
+
+  To "propagate" a work means to do anything with it that, without
+permission, would make you directly or secondarily liable for
+infringement under applicable copyright law, except executing it on a
+computer or modifying a private copy.  Propagation includes copying,
+distribution (with or without modification), making available to the
+public, and in some countries other activities as well.
+
+  To "convey" a work means any kind of propagation that enables other
+parties to make or receive copies.  Mere interaction with a user through
+a computer network, with no transfer of a copy, is not conveying.
+
+  An interactive user interface displays "Appropriate Legal Notices"
+to the extent that it includes a convenient and prominently visible
+feature that (1) displays an appropriate copyright notice, and (2)
+tells the user that there is no warranty for the work (except to the
+extent that warranties are provided), that licensees may convey the
+work under this License, and how to view a copy of this License.  If
+the interface presents a list of user commands or options, such as a
+menu, a prominent item in the list meets this criterion.
+
+  1. Source Code.
+
+  The "source code" for a work means the preferred form of the work
+for making modifications to it.  "Object code" means any non-source
+form of a work.
+
+  A "Standard Interface" means an interface that either is an official
+standard defined by a recognized standards body, or, in the case of
+interfaces specified for a particular programming language, one that
+is widely used among developers working in that language.
+
+  The "System Libraries" of an executable work include anything, other
+than the work as a whole, that (a) is included in the normal form of
+packaging a Major Component, but which is not part of that Major
+Component, and (b) serves only to enable use of the work with that
+Major Component, or to implement a Standard Interface for which an
+implementation is available to the public in source code form.  A
+"Major Component", in this context, means a major essential component
+(kernel, window system, and so on) of the specific operating system
+(if any) on which the executable work runs, or a compiler used to
+produce the work, or an object code interpreter used to run it.
+
+  The "Corresponding Source" for a work in object code form means all
+the source code needed to generate, install, and (for an executable
+work) run the object code and to modify the work, including scripts to
+control those activities.  However, it does not include the work's
+System Libraries, or general-purpose tools or generally available free
+programs which are used unmodified in performing those activities but
+which are not part of the work.  For example, Corresponding Source
+includes interface definition files associated with source files for
+the work, and the source code for shared libraries and dynamically
+linked subprograms that the work is specifically designed to require,
+such as by intimate data communication or control flow between those
+subprograms and other parts of the work.
+
+  The Corresponding Source need not include anything that users
+can regenerate automatically from other parts of the Corresponding
+Source.
+
+  The Corresponding Source for a work in source code form is that
+same work.
+
+  2. Basic Permissions.
+
+  All rights granted under this License are granted for the term of
+copyright on the Program, and are irrevocable provided the stated
+conditions are met.  This License explicitly affirms your unlimited
+permission to run the unmodified Program.  The output from running a
+covered work is covered by this License only if the output, given its
+content, constitutes a covered work.  This License acknowledges your
+rights of fair use or other equivalent, as provided by copyright law.
+
+  You may make, run and propagate covered works that you do not
+convey, without conditions so long as your license otherwise remains
+in force.  You may convey covered works to others for the sole purpose
+of having them make modifications exclusively for you, or provide you
+with facilities for running those works, provided that you comply with
+the terms of this License in conveying all material for which you do
+not control copyright.  Those thus making or running the covered works
+for you must do so exclusively on your behalf, under your direction
+and control, on terms that prohibit them from making any copies of
+your copyrighted material outside their relationship with you.
+
+  Conveying under any other circumstances is permitted solely under
+the conditions stated below.  Sublicensing is not allowed; section 10
+makes it unnecessary.
+
+  3. Protecting Users' Legal Rights From Anti-Circumvention Law.
+
+  No covered work shall be deemed part of an effective technological
+measure under any applicable law fulfilling obligations under article
+11 of the WIPO copyright treaty adopted on 20 December 1996, or
+similar laws prohibiting or restricting circumvention of such
+measures.
+
+  When you convey a covered work, you waive any legal power to forbid
+circumvention of technological measures to the extent such circumvention
+is effected by exercising rights under this License with respect to
+the covered work, and you disclaim any intention to limit operation or
+modification of the work as a means of enforcing, against the work's
+users, your or third parties' legal rights to forbid circumvention of
+technological measures.
+
+  4. Conveying Verbatim Copies.
+
+  You may convey verbatim copies of the Program's source code as you
+receive it, in any medium, provided that you conspicuously and
+appropriately publish on each copy an appropriate copyright notice;
+keep intact all notices stating that this License and any
+non-permissive terms added in accord with section 7 apply to the code;
+keep intact all notices of the absence of any warranty; and give all
+recipients a copy of this License along with the Program.
+
+  You may charge any price or no price for each copy that you convey,
+and you may offer support or warranty protection for a fee.
+
+  5. Conveying Modified Source Versions.
+
+  You may convey a work based on the Program, or the modifications to
+produce it from the Program, in the form of source code under the
+terms of section 4, provided that you also meet all of these conditions:
+
+    a) The work must carry prominent notices stating that you modified
+    it, and giving a relevant date.
+
+    b) The work must carry prominent notices stating that it is
+    released under this License and any conditions added under section
+    7.  This requirement modifies the requirement in section 4 to
+    "keep intact all notices".
+
+    c) You must license the entire work, as a whole, under this
+    License to anyone who comes into possession of a copy.  This
+    License will therefore apply, along with any applicable section 7
+    additional terms, to the whole of the work, and all its parts,
+    regardless of how they are packaged.  This License gives no
+    permission to license the work in any other way, but it does not
+    invalidate such permission if you have separately received it.
+
+    d) If the work has interactive user interfaces, each must display
+    Appropriate Legal Notices; however, if the Program has interactive
+    interfaces that do not display Appropriate Legal Notices, your
+    work need not make them do so.
+
+  A compilation of a covered work with other separate and independent
+works, which are not by their nature extensions of the covered work,
+and which are not combined with it such as to form a larger program,
+in or on a volume of a storage or distribution medium, is called an
+"aggregate" if the compilation and its resulting copyright are not
+used to limit the access or legal rights of the compilation's users
+beyond what the individual works permit.  Inclusion of a covered work
+in an aggregate does not cause this License to apply to the other
+parts of the aggregate.
+
+  6. Conveying Non-Source Forms.
+
+  You may convey a covered work in object code form under the terms
+of sections 4 and 5, provided that you also convey the
+machine-readable Corresponding Source under the terms of this License,
+in one of these ways:
+
+    a) Convey the object code in, or embodied in, a physical product
+    (including a physical distribution medium), accompanied by the
+    Corresponding Source fixed on a durable physical medium
+    customarily used for software interchange.
+
+    b) Convey the object code in, or embodied in, a physical product
+    (including a physical distribution medium), accompanied by a
+    written offer, valid for at least three years and valid for as
+    long as you offer spare parts or customer support for that product
+    model, to give anyone who possesses the object code either (1) a
+    copy of the Corresponding Source for all the software in the
+    product that is covered by this License, on a durable physical
+    medium customarily used for software interchange, for a price no
+    more than your reasonable cost of physically performing this
+    conveying of source, or (2) access to copy the
+    Corresponding Source from a network server at no charge.
+
+    c) Convey individual copies of the object code with a copy of the
+    written offer to provide the Corresponding Source.  This
+    alternative is allowed only occasionally and noncommercially, and
+    only if you received the object code with such an offer, in accord
+    with subsection 6b.
+
+    d) Convey the object code by offering access from a designated
+    place (gratis or for a charge), and offer equivalent access to the
+    Corresponding Source in the same way through the same place at no
+    further charge.  You need not require recipients to copy the
+    Corresponding Source along with the object code.  If the place to
+    copy the object code is a network server, the Corresponding Source
+    may be on a different server (operated by you or a third party)
+    that supports equivalent copying facilities, provided you maintain
+    clear directions next to the object code saying where to find the
+    Corresponding Source.  Regardless of what server hosts the
+    Corresponding Source, you remain obligated to ensure that it is
+    available for as long as needed to satisfy these requirements.
+
+    e) Convey the object code using peer-to-peer transmission, provided
+    you inform other peers where the object code and Corresponding
+    Source of the work are being offered to the general public at no
+    charge under subsection 6d.
+
+  A separable portion of the object code, whose source code is excluded
+from the Corresponding Source as a System Library, need not be
+included in conveying the object code work.
+
+  A "User Product" is either (1) a "consumer product", which means any
+tangible personal property which is normally used for personal, family,
+or household purposes, or (2) anything designed or sold for incorporation
+into a dwelling.  In determining whether a product is a consumer product,
+doubtful cases shall be resolved in favor of coverage.  For a particular
+product received by a particular user, "normally used" refers to a
+typical or common use of that class of product, regardless of the status
+of the particular user or of the way in which the particular user
+actually uses, or expects or is expected to use, the product.  A product
+is a consumer product regardless of whether the product has substantial
+commercial, industrial or non-consumer uses, unless such uses represent
+the only significant mode of use of the product.
+
+  "Installation Information" for a User Product means any methods,
+procedures, authorization keys, or other information required to install
+and execute modified versions of a covered work in that User Product from
+a modified version of its Corresponding Source.  The information must
+suffice to ensure that the continued functioning of the modified object
+code is in no case prevented or interfered with solely because
+modification has been made.
+
+  If you convey an object code work under this section in, or with, or
+specifically for use in, a User Product, and the conveying occurs as
+part of a transaction in which the right of possession and use of the
+User Product is transferred to the recipient in perpetuity or for a
+fixed term (regardless of how the transaction is characterized), the
+Corresponding Source conveyed under this section must be accompanied
+by the Installation Information.  But this requirement does not apply
+if neither you nor any third party retains the ability to install
+modified object code on the User Product (for example, the work has
+been installed in ROM).
+
+  The requirement to provide Installation Information does not include a
+requirement to continue to provide support service, warranty, or updates
+for a work that has been modified or installed by the recipient, or for
+the User Product in which it has been modified or installed.  Access to a
+network may be denied when the modification itself materially and
+adversely affects the operation of the network or violates the rules and
+protocols for communication across the network.
+
+  Corresponding Source conveyed, and Installation Information provided,
+in accord with this section must be in a format that is publicly
+documented (and with an implementation available to the public in
+source code form), and must require no special password or key for
+unpacking, reading or copying.
+
+  7. Additional Terms.
+
+  "Additional permissions" are terms that supplement the terms of this
+License by making exceptions from one or more of its conditions.
+Additional permissions that are applicable to the entire Program shall
+be treated as though they were included in this License, to the extent
+that they are valid under applicable law.  If additional permissions
+apply only to part of the Program, that part may be used separately
+under those permissions, but the entire Program remains governed by
+this License without regard to the additional permissions.
+
+  When you convey a copy of a covered work, you may at your option
+remove any additional permissions from that copy, or from any part of
+it.  (Additional permissions may be written to require their own
+removal in certain cases when you modify the work.)  You may place
+additional permissions on material, added by you to a covered work,
+for which you have or can give appropriate copyright permission.
+
+  Notwithstanding any other provision of this License, for material you
+add to a covered work, you may (if authorized by the copyright holders of
+that material) supplement the terms of this License with terms:
+
+    a) Disclaiming warranty or limiting liability differently from the
+    terms of sections 15 and 16 of this License; or
+
+    b) Requiring preservation of specified reasonable legal notices or
+    author attributions in that material or in the Appropriate Legal
+    Notices displayed by works containing it; or
+
+    c) Prohibiting misrepresentation of the origin of that material, or
+    requiring that modified versions of such material be marked in
+    reasonable ways as different from the original version; or
+
+    d) Limiting the use for publicity purposes of names of licensors or
+    authors of the material; or
+
+    e) Declining to grant rights under trademark law for use of some
+    trade names, trademarks, or service marks; or
+
+    f) Requiring indemnification of licensors and authors of that
+    material by anyone who conveys the material (or modified versions of
+    it) with contractual assumptions of liability to the recipient, for
+    any liability that these contractual assumptions directly impose on
+    those licensors and authors.
+
+  All other non-permissive additional terms are considered "further
+restrictions" within the meaning of section 10.  If the Program as you
+received it, or any part of it, contains a notice stating that it is
+governed by this License along with a term that is a further
+restriction, you may remove that term.  If a license document contains
+a further restriction but permits relicensing or conveying under this
+License, you may add to a covered work material governed by the terms
+of that license document, provided that the further restriction does
+not survive such relicensing or conveying.
+
+  If you add terms to a covered work in accord with this section, you
+must place, in the relevant source files, a statement of the
+additional terms that apply to those files, or a notice indicating
+where to find the applicable terms.
+
+  Additional terms, permissive or non-permissive, may be stated in the
+form of a separately written license, or stated as exceptions;
+the above requirements apply either way.
+
+  8. Termination.
+
+  You may not propagate or modify a covered work except as expressly
+provided under this License.  Any attempt otherwise to propagate or
+modify it is void, and will automatically terminate your rights under
+this License (including any patent licenses granted under the third
+paragraph of section 11).
+
+  However, if you cease all violation of this License, then your
+license from a particular copyright holder is reinstated (a)
+provisionally, unless and until the copyright holder explicitly and
+finally terminates your license, and (b) permanently, if the copyright
+holder fails to notify you of the violation by some reasonable means
+prior to 60 days after the cessation.
+
+  Moreover, your license from a particular copyright holder is
+reinstated permanently if the copyright holder notifies you of the
+violation by some reasonable means, this is the first time you have
+received notice of violation of this License (for any work) from that
+copyright holder, and you cure the violation prior to 30 days after
+your receipt of the notice.
+
+  Termination of your rights under this section does not terminate the
+licenses of parties who have received copies or rights from you under
+this License.  If your rights have been terminated and not permanently
+reinstated, you do not qualify to receive new licenses for the same
+material under section 10.
+
+  9. Acceptance Not Required for Having Copies.
+
+  You are not required to accept this License in order to receive or
+run a copy of the Program.  Ancillary propagation of a covered work
+occurring solely as a consequence of using peer-to-peer transmission
+to receive a copy likewise does not require acceptance.  However,
+nothing other than this License grants you permission to propagate or
+modify any covered work.  These actions infringe copyright if you do
+not accept this License.  Therefore, by modifying or propagating a
+covered work, you indicate your acceptance of this License to do so.
+
+  10. Automatic Licensing of Downstream Recipients.
+
+  Each time you convey a covered work, the recipient automatically
+receives a license from the original licensors, to run, modify and
+propagate that work, subject to this License.  You are not responsible
+for enforcing compliance by third parties with this License.
+
+  An "entity transaction" is a transaction transferring control of an
+organization, or substantially all assets of one, or subdividing an
+organization, or merging organizations.  If propagation of a covered
+work results from an entity transaction, each party to that
+transaction who receives a copy of the work also receives whatever
+licenses to the work the party's predecessor in interest had or could
+give under the previous paragraph, plus a right to possession of the
+Corresponding Source of the work from the predecessor in interest, if
+the predecessor has it or can get it with reasonable efforts.
+
+  You may not impose any further restrictions on the exercise of the
+rights granted or affirmed under this License.  For example, you may
+not impose a license fee, royalty, or other charge for exercise of
+rights granted under this License, and you may not initiate litigation
+(including a cross-claim or counterclaim in a lawsuit) alleging that
+any patent claim is infringed by making, using, selling, offering for
+sale, or importing the Program or any portion of it.
+
+  11. Patents.
+
+  A "contributor" is a copyright holder who authorizes use under this
+License of the Program or a work on which the Program is based.  The
+work thus licensed is called the contributor's "contributor version".
+
+  A contributor's "essential patent claims" are all patent claims
+owned or controlled by the contributor, whether already acquired or
+hereafter acquired, that would be infringed by some manner, permitted
+by this License, of making, using, or selling its contributor version,
+but do not include claims that would be infringed only as a
+consequence of further modification of the contributor version.  For
+purposes of this definition, "control" includes the right to grant
+patent sublicenses in a manner consistent with the requirements of
+this License.
+
+  Each contributor grants you a non-exclusive, worldwide, royalty-free
+patent license under the contributor's essential patent claims, to
+make, use, sell, offer for sale, import and otherwise run, modify and
+propagate the contents of its contributor version.
+
+  In the following three paragraphs, a "patent license" is any express
+agreement or commitment, however denominated, not to enforce a patent
+(such as an express permission to practice a patent or covenant not to
+sue for patent infringement).  To "grant" such a patent license to a
+party means to make such an agreement or commitment not to enforce a
+patent against the party.
+
+  If you convey a covered work, knowingly relying on a patent license,
+and the Corresponding Source of the work is not available for anyone
+to copy, free of charge and under the terms of this License, through a
+publicly available network server or other readily accessible means,
+then you must either (1) cause the Corresponding Source to be so
+available, or (2) arrange to deprive yourself of the benefit of the
+patent license for this particular work, or (3) arrange, in a manner
+consistent with the requirements of this License, to extend the patent
+license to downstream recipients.  "Knowingly relying" means you have
+actual knowledge that, but for the patent license, your conveying the
+covered work in a country, or your recipient's use of the covered work
+in a country, would infringe one or more identifiable patents in that
+country that you have reason to believe are valid.
+
+  If, pursuant to or in connection with a single transaction or
+arrangement, you convey, or propagate by procuring conveyance of, a
+covered work, and grant a patent license to some of the parties
+receiving the covered work authorizing them to use, propagate, modify
+or convey a specific copy of the covered work, then the patent license
+you grant is automatically extended to all recipients of the covered
+work and works based on it.
+
+  A patent license is "discriminatory" if it does not include within
+the scope of its coverage, prohibits the exercise of, or is
+conditioned on the non-exercise of one or more of the rights that are
+specifically granted under this License.  You may not convey a covered
+work if you are a party to an arrangement with a third party that is
+in the business of distributing software, under which you make payment
+to the third party based on the extent of your activity of conveying
+the work, and under which the third party grants, to any of the
+parties who would receive the covered work from you, a discriminatory
+patent license (a) in connection with copies of the covered work
+conveyed by you (or copies made from those copies), or (b) primarily
+for and in connection with specific products or compilations that
+contain the covered work, unless you entered into that arrangement,
+or that patent license was granted, prior to 28 March 2007.
+
+  Nothing in this License shall be construed as excluding or limiting
+any implied license or other defenses to infringement that may
+otherwise be available to you under applicable patent law.
+
+  12. No Surrender of Others' Freedom.
+
+  If conditions are imposed on you (whether by court order, agreement or
+otherwise) that contradict the conditions of this License, they do not
+excuse you from the conditions of this License.  If you cannot convey a
+covered work so as to satisfy simultaneously your obligations under this
+License and any other pertinent obligations, then as a consequence you may
+not convey it at all.  For example, if you agree to terms that obligate you
+to collect a royalty for further conveying from those to whom you convey
+the Program, the only way you could satisfy both those terms and this
+License would be to refrain entirely from conveying the Program.
+
+  13. Remote Network Interaction; Use with the GNU General Public License.
+
+  Notwithstanding any other provision of this License, if you modify the
+Program, your modified version must prominently offer all users
+interacting with it remotely through a computer network (if your version
+supports such interaction) an opportunity to receive the Corresponding
+Source of your version by providing access to the Corresponding Source
+from a network server at no charge, through some standard or customary
+means of facilitating copying of software.  This Corresponding Source
+shall include the Corresponding Source for any work covered by version 3
+of the GNU General Public License that is incorporated pursuant to the
+following paragraph.
+
+  Notwithstanding any other provision of this License, you have
+permission to link or combine any covered work with a work licensed
+under version 3 of the GNU General Public License into a single
+combined work, and to convey the resulting work.  The terms of this
+License will continue to apply to the part which is the covered work,
+but the work with which it is combined will remain governed by version
+3 of the GNU General Public License.
+
+  14. Revised Versions of this License.
+
+  The Free Software Foundation may publish revised and/or new versions of
+the GNU Affero General Public License from time to time.  Such new versions
+will be similar in spirit to the present version, but may differ in detail to
+address new problems or concerns.
+
+  Each version is given a distinguishing version number.  If the
+Program specifies that a certain numbered version of the GNU Affero General
+Public License "or any later version" applies to it, you have the
+option of following the terms and conditions either of that numbered
+version or of any later version published by the Free Software
+Foundation.  If the Program does not specify a version number of the
+GNU Affero General Public License, you may choose any version ever published
+by the Free Software Foundation.
+
+  If the Program specifies that a proxy can decide which future
+versions of the GNU Affero General Public License can be used, that proxy's
+public statement of acceptance of a version permanently authorizes you
+to choose that version for the Program.
+
+  Later license versions may give you additional or different
+permissions.  However, no additional obligations are imposed on any
+author or copyright holder as a result of your choosing to follow a
+later version.
+
+  15. Disclaimer of Warranty.
+
+  THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
+APPLICABLE LAW.  EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
+HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
+OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
+THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+PURPOSE.  THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
+IS WITH YOU.  SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
+ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
+
+  16. Limitation of Liability.
+
+  IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
+WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
+THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
+GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
+USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
+DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
+PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
+EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
+SUCH DAMAGES.
+
+  17. Interpretation of Sections 15 and 16.
+
+  If the disclaimer of warranty and limitation of liability provided
+above cannot be given local legal effect according to their terms,
+reviewing courts shall apply local law that most closely approximates
+an absolute waiver of all civil liability in connection with the
+Program, unless a warranty or assumption of liability accompanies a
+copy of the Program in return for a fee.
+
+                     END OF TERMS AND CONDITIONS
+
+            How to Apply These Terms to Your New Programs
+
+  If you develop a new program, and you want it to be of the greatest
+possible use to the public, the best way to achieve this is to make it
+free software which everyone can redistribute and change under these terms.
+
+  To do so, attach the following notices to the program.  It is safest
+to attach them to the start of each source file to most effectively
+state the exclusion of warranty; and each file should have at least
+the "copyright" line and a pointer to where the full notice is found.
+
+    <one line to give the program's name and a brief idea of what it does.>
+    Copyright (C) <year>  <name of author>
+
+    This program is free software: you can redistribute it and/or modify
+    it under the terms of the GNU Affero General Public License as published
+    by the Free Software Foundation, either version 3 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero General Public License
+    along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+Also add information on how to contact you by electronic and paper mail.
+
+  If your software can interact with users remotely through a computer
+network, you should also make sure that it provides a way for users to
+get its source.  For example, if your program is a web application, its
+interface could display a "Source" link that leads users to an archive
+of the code.  There are many ways you could offer source, and different
+solutions will be better for different programs; see section 13 for the
+specific requirements.
+
+  You should also get your employer (if you work as a programmer) or school,
+if any, to sign a "copyright disclaimer" for the program, if necessary.
+For more information on this, and how to apply and follow the GNU AGPL, see
+<https://www.gnu.org/licenses/>.

README.md

@@ -1,13 +1,114 @@
----
-title: FaceSwap
-emoji: 🏃
-colorFrom: blue
-colorTo: blue
-sdk: gradio
-sdk_version: 3.44.3
-app_file: app.py
-pinned: false
-license: openrail
----
-
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# roop-unleashed
+
+[Changelog](#changelog) • [Installation](#installation) • [Usage](#usage) • [Example](#example) • [FAQ](#faq)
+
+
+Uncensored Deepfakes for images and videos without training and an easy-to-use GUI.
+
+
+![Screen](docs/screenshot.png)
+
+
+### Features
+
+- Platform-independant Browser GUI
+- Selection of multiple input/output faces in one go
+- Many different swapping modes, first detected, face selections, by gender
+- Batch processing of images/videos
+- Masking of face occluders using text prompts
+- Optional Face Restoration using different enhancers
+- Preview swapping from different video frames
+- Live Fake Cam using your webcam
+- Extras Tab for cutting videos etc.
+- Settings - storing configuration for next session
+- Theme Support
+
+and lots more...
+
+
+## Disclaimer
+
+This project is for technical and academic use only.
+Users of this software are expected to use this software responsibly while abiding the local law. If a face of a real person is being used, users are suggested to get consent from the concerned person and clearly mention that it is a deepfake when posting content online. Developers of this software will not be responsible for actions of end-users.
+**Please do not apply it to illegal and unethical scenarios.**
+
+In the event of violation of the legal and ethical requirements of the user's country or region, this code repository is exempt from liability
+
+### Installation
+
+> For Windows, you need to download and install [Visual Studio](https://visualstudio.microsoft.com/de/downloads/) (in theory build-tools might work too but in my experience so far they don't). During the install, make sure to include the C++ package.
+
+Besides that, just use the 1-click installer in releases. This will download and install everything
+in a handy conda environment. This not only installs the application but also runs it, once installed.
+
+For other OS or if you know what you're doing:
+
+- `git clone https://github.com/C0untFloyd/roop-unleashed`
+- preferably create a venv or conda environment
+- `cd roop-unleashed`
+- `pip install -r requirements.txt`
+
+Depending on your available GPU there are additional packages you need to install. Here are the instructions from the original roop page:
+
+[Using GPU Acceleration](https://github.com/s0md3v/roop/wiki/2.-Acceleration)
+
+The used GPU Provider is configured in the settings tab, no need to use cmdline arguments any more. Default is CUDA (for NVIDIA). If you change it, please restart roop-unleashed completely to allow for model reloading.
+
+For Video face-swapping you also need to have ffmpeg properly installed (having it in your PATH Env). The windows installer tries to do this automatically.  
+
+
+
+### Usage
+
+- Windows: run the `windows_run.bat` from the Installer.
+- Linux: `python run.py`
+
+<a target="_blank" href="https://colab.research.google.com/github/C0untFloyd/roop-unleashed/blob/main/roop-unleashed.ipynb">
+  <img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"/>
+</a>
+  
+
+Additional commandline arguments are currently unsupported and settings should be done via the UI.
+
+> Note: When you run this program for the first time, it will download some models roughly ~2Gb in size.
+
+
+### Example
+
+*Coming soon*
+
+
+
+### Changelog
+
+**11.8.2023** v2.7.0
+
+Initial Gradio Version - old TkInter Version now deprecated
+
+- Re-added unified padding to face enhancers
+- Fixed DMDNet for all resolutions
+- Selecting target face now automatically switches swapping mode to selected
+- GPU providers are correctly set using the GUI (needs restart currently)
+- Local output folder can be opened from page
+- Unfinished extras functions disabled for now
+- Installer checks out specific commit, allowing to go back to first install
+- Updated readme for new gradio version
+- Updated Colab
+
+
+# Acknowledgements
+
+Lots of ideas, code or pre-trained models used from the following projects:
+
+https://github.com/deepinsight/insightface
+https://github.com/s0md3v/roop
+https://github.com/AUTOMATIC1111/stable-diffusion-webui
+https://github.com/Hillobar/Rope
+https://github.com/janvarev/chain-img-processor
+https://github.com/TencentARC/GFPGAN   
+https://github.com/kadirnar/codeformer-pip
+https://github.com/csxmli2016/DMDNet
+
+
+Thanks to all developers!
+

chain_img_processor/image.py

@@ -0,0 +1,222 @@
+import numpy as np
+import cv2
+from jaa import JaaCore
+from roop.utilities import get_device
+
+
+from typing import Any
+
+version = "4.0.0"
+
+class ChainImgProcessor(JaaCore):
+
+    def __init__(self):
+        JaaCore.__init__(self)
+
+        self.processors:dict = {
+        }
+
+        self.processors_objects:dict[str,list[ChainImgPlugin]] = {}
+
+        self.default_chain = ""
+        self.init_on_start = ""
+
+        self.inited_processors = []
+
+        self.is_demo_row_render = False
+
+    def process_plugin_manifest(self, modname, manifest):
+        # adding processors from plugin manifest
+        if "img_processor" in manifest:  # process commands
+            for cmd in manifest["img_processor"].keys():
+                self.processors[cmd] = manifest["img_processor"][cmd]
+
+        return manifest
+
+    def init_with_plugins(self):
+        self.init_plugins(["core"])
+        self.display_init_info()
+
+        #self.init_translator_engine(self.default_translator)
+        init_on_start_arr = self.init_on_start.split(",")
+        for proc_id in init_on_start_arr:
+            self.init_processor(proc_id)
+
+    def run_chain(self, img, params:dict[str,Any] = None, chain:str = None, thread_index:int = 0):
+        if chain is None:
+            chain = self.default_chain
+        if params is None:
+            params = {}
+        params["_thread_index"] = thread_index
+        chain_ar = chain.split(",")
+        # init all not inited processors first
+        for proc_id in chain_ar:
+            if proc_id != "":
+                if not proc_id in self.inited_processors:
+                    self.init_processor(proc_id)
+
+
+
+        # run processing
+        if self.is_demo_row_render:
+            import cv2
+            import numpy as np
+            height, width, channels = img.shape
+            img_blank = np.zeros((height+30, width*(1+len(chain_ar)), 3), dtype=np.uint8)
+            img_blank.fill(255)
+
+            y = 30
+            x = 0
+            img_blank[y:y + height, x:x + width] = img
+
+            # Set the font scale and thickness
+            font_scale = 1
+            thickness = 2
+
+            # Set the font face to a monospace font
+            font_face = cv2.FONT_HERSHEY_SIMPLEX
+
+            cv2.putText(img_blank, "original", (x+4, y-7), font_face, font_scale, (0, 0, 0), thickness)
+
+
+        i = 0
+        for proc_id in chain_ar:
+            i += 1
+            if proc_id != "":
+                #img = self.processors[proc_id][1](self, img, params) # params can be modified inside
+                y = 30
+                img = self.processors_objects[proc_id][thread_index].process(img,params)
+                if self.is_demo_row_render:
+                    x = width*i
+                    img_blank[y:y + height, x:x + width] = img
+                    cv2.putText(img_blank, proc_id, (x + 4, y - 7), font_face, font_scale, (0, 0, 0), thickness)
+
+        if self.is_demo_row_render:
+            return img_blank, params
+
+        return img, params
+
+    # ---------------- init translation stuff ----------------
+    def fill_processors_for_thread_chains(self, threads:int = 1, chain:str = None):
+        if chain is None:
+            chain = self.default_chain
+
+        chain_ar = chain.split(",")
+        # init all not initialized processors first
+        for processor_id in chain_ar:
+            if processor_id != "":
+                if self.processors_objects.get(processor_id) is None:
+                    self.processors_objects[processor_id] = []
+                while len(self.processors_objects[processor_id]) < threads:
+                    self.add_processor_to_list(processor_id)
+
+    def add_processor_to_list(self, processor_id: str):
+        obj = self.processors[processor_id](self)
+        obj.init_plugin()
+        if self.processors_objects.get(processor_id) is None:
+            self.processors_objects[processor_id] = []
+        self.processors_objects[processor_id].append(obj)
+    def init_processor(self, processor_id: str):
+        if processor_id == "": # blank line case
+            return
+
+        if processor_id in self.inited_processors:
+            return
+
+        try:
+            if self.verbose:
+                self.print_blue("TRY: init processor plugin '{0}'...".format(processor_id))
+            self.add_processor_to_list(processor_id)
+            self.inited_processors.append(processor_id)
+            if self.verbose:
+                self.print_blue("SUCCESS: '{0}' initialized!".format(processor_id))
+
+        except Exception as e:
+            self.print_error("Error init processor plugin {0}...".format(processor_id), e)
+
+    # ------------ formatting stuff -------------------
+    def display_init_info(self):
+        if self.verbose:
+            print("ChainImgProcessor v{0}:".format(version))
+            self.format_print_key_list("processors:", self.processors.keys())
+
+    def format_print_key_list(self, key:str, value:list):
+        print(key+": ".join(value))
+
+    def print_error(self,err_txt,e:Exception = None):
+        print(err_txt,"red")
+        # if e != None:
+        #     cprint(e,"red")
+        import traceback
+        traceback.print_exc()
+
+    def print_red(self,txt):
+        print(txt)
+
+    def print_blue(self, txt):
+        print(txt)
+
+class ChainImgPlugin:
+
+    device = 'cpu'
+
+    def __init__(self, core: ChainImgProcessor):
+        self.core = core
+        self.device = get_device()
+
+    def init_plugin(self): # here you can init something. Called once
+        pass
+    def process(self, img, params:dict): # process img. Called multiple
+        return img
+    
+    def unload(self):
+        pass
+
+
+    def cutout(self, frame, start_x, start_y, end_x, end_y, padding_factor):
+        padding_x = int((end_x - start_x) * padding_factor)
+        padding_y = int((end_y - start_y) * padding_factor)
+
+        start_x = max(0, start_x - padding_x)
+        start_y = max(0, start_y - padding_y)
+        end_x = min(frame.shape[1], end_x + padding_x)
+        end_y = min(frame.shape[0], end_y + padding_y)
+        return frame[start_y:end_y, start_x:end_x], start_x, start_y, end_x, end_y
+    
+    def paste_into(self, clip, frame, start_x, start_y, end_x, end_y, smooth):
+        if smooth:
+            smallest = min(clip.shape[0], clip.shape[1])
+            mask_border = smallest // 12
+            if mask_border > 4:
+                img_white = np.full((clip.shape[0], clip.shape[1]), 0, dtype=float)
+                # img_white = cv2.warpAffine(img_white, mat_rev, img_shape)
+                # img_white[img_white > 20] = 255
+                img_white = cv2.rectangle(img_white, (mask_border, mask_border), 
+                                        (img_white.shape[1] - mask_border, img_white.shape[0]-mask_border), (255, 255, 255), -1)    
+                img_mask = img_white
+                t1 = mask_border * 2
+                kernel = np.ones((t1, t1), np.uint8)
+                img_mask = cv2.erode(img_mask, kernel, iterations=2)
+                t1 = mask_border
+                kernel_size = (t1, t1)
+                blur_size = tuple(2 * j + 1 for j in kernel_size)
+                img_mask = cv2.GaussianBlur(img_mask, blur_size, 0)
+                img_mask /= 255
+                img_mask = np.reshape(img_mask, [img_mask.shape[0], img_mask.shape[1], 1])
+                frame_clip = frame[start_y:end_y, start_x:end_x]
+                clip = img_mask * clip + (1 - img_mask) * frame_clip
+                
+        frame[start_y:end_y, start_x:end_x] = clip
+        return frame
+
+    
+
+
+_img_processor:ChainImgProcessor = None
+def get_single_image_processor() -> ChainImgProcessor:
+    global _img_processor
+    if _img_processor is None:
+        _img_processor = ChainImgProcessor()
+        _img_processor.init_with_plugins()
+    return _img_processor
+

clip/__init__.py

@@ -0,0 +1 @@
+from .clip import *

clip/bpe_simple_vocab_16e6.txt.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:924691ac288e54409236115652ad4aa250f48203de50a9e4722a6ecd48d6804a
+size 1356917

clip/clip.py

@@ -0,0 +1,245 @@
+import hashlib
+import os
+import urllib
+import warnings
+from typing import Any, Union, List
+from pkg_resources import packaging
+
+import torch
+from PIL import Image
+from torchvision.transforms import Compose, Resize, CenterCrop, ToTensor, Normalize
+from tqdm import tqdm
+
+from .model import build_model
+from .simple_tokenizer import SimpleTokenizer as _Tokenizer
+
+try:
+    from torchvision.transforms import InterpolationMode
+    BICUBIC = InterpolationMode.BICUBIC
+except ImportError:
+    BICUBIC = Image.BICUBIC
+
+
+if packaging.version.parse(torch.__version__) < packaging.version.parse("1.7.1"):
+    warnings.warn("PyTorch version 1.7.1 or higher is recommended")
+
+
+__all__ = ["available_models", "load", "tokenize"]
+_tokenizer = _Tokenizer()
+
+_MODELS = {
+    "RN50": "https://openaipublic.azureedge.net/clip/models/afeb0e10f9e5a86da6080e35cf09123aca3b358a0c3e3b6c78a7b63bc04b6762/RN50.pt",
+    "RN101": "https://openaipublic.azureedge.net/clip/models/8fa8567bab74a42d41c5915025a8e4538c3bdbe8804a470a72f30b0d94fab599/RN101.pt",
+    "RN50x4": "https://openaipublic.azureedge.net/clip/models/7e526bd135e493cef0776de27d5f42653e6b4c8bf9e0f653bb11773263205fdd/RN50x4.pt",
+    "RN50x16": "https://openaipublic.azureedge.net/clip/models/52378b407f34354e150460fe41077663dd5b39c54cd0bfd2b27167a4a06ec9aa/RN50x16.pt",
+    "RN50x64": "https://openaipublic.azureedge.net/clip/models/be1cfb55d75a9666199fb2206c106743da0f6468c9d327f3e0d0a543a9919d9c/RN50x64.pt",
+    "ViT-B/32": "https://openaipublic.azureedge.net/clip/models/40d365715913c9da98579312b702a82c18be219cc2a73407c4526f58eba950af/ViT-B-32.pt",
+    "ViT-B/16": "https://openaipublic.azureedge.net/clip/models/5806e77cd80f8b59890b7e101eabd078d9fb84e6937f9e85e4ecb61988df416f/ViT-B-16.pt",
+    "ViT-L/14": "https://openaipublic.azureedge.net/clip/models/b8cca3fd41ae0c99ba7e8951adf17d267cdb84cd88be6f7c2e0eca1737a03836/ViT-L-14.pt",
+    "ViT-L/14@336px": "https://openaipublic.azureedge.net/clip/models/3035c92b350959924f9f00213499208652fc7ea050643e8b385c2dac08641f02/ViT-L-14-336px.pt",
+}
+
+
+def _download(url: str, root: str):
+    os.makedirs(root, exist_ok=True)
+    filename = os.path.basename(url)
+
+    expected_sha256 = url.split("/")[-2]
+    download_target = os.path.join(root, filename)
+
+    if os.path.exists(download_target) and not os.path.isfile(download_target):
+        raise RuntimeError(f"{download_target} exists and is not a regular file")
+
+    if os.path.isfile(download_target):
+        if hashlib.sha256(open(download_target, "rb").read()).hexdigest() == expected_sha256:
+            return download_target
+        else:
+            warnings.warn(f"{download_target} exists, but the SHA256 checksum does not match; re-downloading the file")
+
+    with urllib.request.urlopen(url) as source, open(download_target, "wb") as output:
+        with tqdm(total=int(source.info().get("Content-Length")), ncols=80, unit='iB', unit_scale=True, unit_divisor=1024) as loop:
+            while True:
+                buffer = source.read(8192)
+                if not buffer:
+                    break
+
+                output.write(buffer)
+                loop.update(len(buffer))
+
+    if hashlib.sha256(open(download_target, "rb").read()).hexdigest() != expected_sha256:
+        raise RuntimeError("Model has been downloaded but the SHA256 checksum does not not match")
+
+    return download_target
+
+
+def _convert_image_to_rgb(image):
+    return image.convert("RGB")
+
+
+def _transform(n_px):
+    return Compose([
+        Resize(n_px, interpolation=BICUBIC),
+        CenterCrop(n_px),
+        _convert_image_to_rgb,
+        ToTensor(),
+        Normalize((0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711)),
+    ])
+
+
+def available_models() -> List[str]:
+    """Returns the names of available CLIP models"""
+    return list(_MODELS.keys())
+
+
+def load(name: str, device: Union[str, torch.device] = "cuda" if torch.cuda.is_available() else "cpu", jit: bool = False, download_root: str = None):
+    """Load a CLIP model
+
+    Parameters
+    ----------
+    name : str
+        A model name listed by `clip.available_models()`, or the path to a model checkpoint containing the state_dict
+
+    device : Union[str, torch.device]
+        The device to put the loaded model
+
+    jit : bool
+        Whether to load the optimized JIT model or more hackable non-JIT model (default).
+
+    download_root: str
+        path to download the model files; by default, it uses "~/.cache/clip"
+
+    Returns
+    -------
+    model : torch.nn.Module
+        The CLIP model
+
+    preprocess : Callable[[PIL.Image], torch.Tensor]
+        A torchvision transform that converts a PIL image into a tensor that the returned model can take as its input
+    """
+    if name in _MODELS:
+        model_path = _download(_MODELS[name], download_root or os.path.expanduser("~/.cache/clip"))
+    elif os.path.isfile(name):
+        model_path = name
+    else:
+        raise RuntimeError(f"Model {name} not found; available models = {available_models()}")
+
+    with open(model_path, 'rb') as opened_file:
+        try:
+            # loading JIT archive
+            model = torch.jit.load(opened_file, map_location=device if jit else "cpu").eval()
+            state_dict = None
+        except RuntimeError:
+            # loading saved state dict
+            if jit:
+                warnings.warn(f"File {model_path} is not a JIT archive. Loading as a state dict instead")
+                jit = False
+            state_dict = torch.load(opened_file, map_location="cpu")
+
+    if not jit:
+        model = build_model(state_dict or model.state_dict()).to(device)
+        if str(device) == "cpu":
+            model.float()
+        return model, _transform(model.visual.input_resolution)
+
+    # patch the device names
+    device_holder = torch.jit.trace(lambda: torch.ones([]).to(torch.device(device)), example_inputs=[])
+    device_node = [n for n in device_holder.graph.findAllNodes("prim::Constant") if "Device" in repr(n)][-1]
+
+    def _node_get(node: torch._C.Node, key: str):
+        """Gets attributes of a node which is polymorphic over return type.
+        
+        From https://github.com/pytorch/pytorch/pull/82628
+        """
+        sel = node.kindOf(key)
+        return getattr(node, sel)(key)
+
+    def patch_device(module):
+        try:
+            graphs = [module.graph] if hasattr(module, "graph") else []
+        except RuntimeError:
+            graphs = []
+
+        if hasattr(module, "forward1"):
+            graphs.append(module.forward1.graph)
+
+        for graph in graphs:
+            for node in graph.findAllNodes("prim::Constant"):
+                if "value" in node.attributeNames() and str(_node_get(node, "value")).startswith("cuda"):
+                    node.copyAttributes(device_node)
+
+    model.apply(patch_device)
+    patch_device(model.encode_image)
+    patch_device(model.encode_text)
+
+    # patch dtype to float32 on CPU
+    if str(device) == "cpu":
+        float_holder = torch.jit.trace(lambda: torch.ones([]).float(), example_inputs=[])
+        float_input = list(float_holder.graph.findNode("aten::to").inputs())[1]
+        float_node = float_input.node()
+
+        def patch_float(module):
+            try:
+                graphs = [module.graph] if hasattr(module, "graph") else []
+            except RuntimeError:
+                graphs = []
+
+            if hasattr(module, "forward1"):
+                graphs.append(module.forward1.graph)
+
+            for graph in graphs:
+                for node in graph.findAllNodes("aten::to"):
+                    inputs = list(node.inputs())
+                    for i in [1, 2]:  # dtype can be the second or third argument to aten::to()
+                        if _node_get(inputs[i].node(), "value") == 5:
+                            inputs[i].node().copyAttributes(float_node)
+
+        model.apply(patch_float)
+        patch_float(model.encode_image)
+        patch_float(model.encode_text)
+
+        model.float()
+
+    return model, _transform(model.input_resolution.item())
+
+
+def tokenize(texts: Union[str, List[str]], context_length: int = 77, truncate: bool = False) -> Union[torch.IntTensor, torch.LongTensor]:
+    """
+    Returns the tokenized representation of given input string(s)
+
+    Parameters
+    ----------
+    texts : Union[str, List[str]]
+        An input string or a list of input strings to tokenize
+
+    context_length : int
+        The context length to use; all CLIP models use 77 as the context length
+
+    truncate: bool
+        Whether to truncate the text in case its encoding is longer than the context length
+
+    Returns
+    -------
+    A two-dimensional tensor containing the resulting tokens, shape = [number of input strings, context_length].
+    We return LongTensor when torch version is <1.8.0, since older index_select requires indices to be long.
+    """
+    if isinstance(texts, str):
+        texts = [texts]
+
+    sot_token = _tokenizer.encoder["<|startoftext|>"]
+    eot_token = _tokenizer.encoder["<|endoftext|>"]
+    all_tokens = [[sot_token] + _tokenizer.encode(text) + [eot_token] for text in texts]
+    if packaging.version.parse(torch.__version__) < packaging.version.parse("1.8.0"):
+        result = torch.zeros(len(all_tokens), context_length, dtype=torch.long)
+    else:
+        result = torch.zeros(len(all_tokens), context_length, dtype=torch.int)
+
+    for i, tokens in enumerate(all_tokens):
+        if len(tokens) > context_length:
+            if truncate:
+                tokens = tokens[:context_length]
+                tokens[-1] = eot_token
+            else:
+                raise RuntimeError(f"Input {texts[i]} is too long for context length {context_length}")
+        result[i, :len(tokens)] = torch.tensor(tokens)
+
+    return result

clip/clipseg.py

@@ -0,0 +1,538 @@
+import math
+from os.path import basename, dirname, join, isfile
+import torch
+from torch import nn
+from torch.nn import functional as nnf
+from torch.nn.modules.activation import ReLU
+
+
+def get_prompt_list(prompt):
+    if prompt == 'plain':
+        return ['{}']    
+    elif prompt == 'fixed':
+        return ['a photo of a {}.']
+    elif prompt == 'shuffle':
+        return ['a photo of a {}.', 'a photograph of a {}.', 'an image of a {}.', '{}.']
+    elif prompt == 'shuffle+':
+        return ['a photo of a {}.', 'a photograph of a {}.', 'an image of a {}.', '{}.',
+                            'a cropped photo of a {}.', 'a good photo of a {}.', 'a photo of one {}.',
+                            'a bad photo of a {}.', 'a photo of the {}.']
+    else:
+        raise ValueError('Invalid value for prompt')        
+
+
+def forward_multihead_attention(x, b, with_aff=False, attn_mask=None):
+    """ 
+    Simplified version of multihead attention (taken from torch source code but without tons of if clauses). 
+    The mlp and layer norm come from CLIP.
+    x: input.
+    b: multihead attention module. 
+    """
+
+    x_ = b.ln_1(x)
+    q, k, v = nnf.linear(x_, b.attn.in_proj_weight, b.attn.in_proj_bias).chunk(3, dim=-1)
+    tgt_len, bsz, embed_dim = q.size()
+
+    head_dim = embed_dim // b.attn.num_heads
+    scaling = float(head_dim) ** -0.5
+
+    q = q.contiguous().view(tgt_len, bsz * b.attn.num_heads, b.attn.head_dim).transpose(0, 1)
+    k = k.contiguous().view(-1, bsz * b.attn.num_heads, b.attn.head_dim).transpose(0, 1)
+    v = v.contiguous().view(-1, bsz * b.attn.num_heads, b.attn.head_dim).transpose(0, 1)
+
+    q = q * scaling
+
+    attn_output_weights = torch.bmm(q, k.transpose(1, 2)) #  n_heads * batch_size, tokens^2, tokens^2
+    if attn_mask is not None:
+
+
+        attn_mask_type, attn_mask = attn_mask
+        n_heads = attn_output_weights.size(0) // attn_mask.size(0)
+        attn_mask = attn_mask.repeat(n_heads, 1)
+        
+        if attn_mask_type == 'cls_token':
+            # the mask only affects similarities compared to the readout-token.
+            attn_output_weights[:, 0, 1:] = attn_output_weights[:, 0, 1:] * attn_mask[None,...]
+            # attn_output_weights[:, 0, 0] = 0*attn_output_weights[:, 0, 0]
+
+        if attn_mask_type == 'all':
+            # print(attn_output_weights.shape, attn_mask[:, None].shape)
+            attn_output_weights[:, 1:, 1:] = attn_output_weights[:, 1:, 1:] * attn_mask[:, None]
+        
+    
+    attn_output_weights = torch.softmax(attn_output_weights, dim=-1)
+
+    attn_output = torch.bmm(attn_output_weights, v)
+    attn_output = attn_output.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim)
+    attn_output = b.attn.out_proj(attn_output)
+
+    x = x + attn_output
+    x = x + b.mlp(b.ln_2(x))
+
+    if with_aff:
+        return x, attn_output_weights
+    else:
+        return x
+
+
+class CLIPDenseBase(nn.Module):
+
+    def __init__(self, version, reduce_cond, reduce_dim, prompt, n_tokens):
+        super().__init__()
+
+        import clip
+
+        # prec = torch.FloatTensor
+        self.clip_model, _ = clip.load(version, device='cpu', jit=False)
+        self.model = self.clip_model.visual
+
+        # if not None, scale conv weights such that we obtain n_tokens.
+        self.n_tokens = n_tokens
+
+        for p in self.clip_model.parameters():
+            p.requires_grad_(False)
+
+        # conditional
+        if reduce_cond is not None:
+            self.reduce_cond = nn.Linear(512, reduce_cond)
+            for p in self.reduce_cond.parameters():
+                p.requires_grad_(False)
+        else:
+            self.reduce_cond = None        
+
+        self.film_mul = nn.Linear(512 if reduce_cond is None else reduce_cond, reduce_dim)
+        self.film_add = nn.Linear(512 if reduce_cond is None else reduce_cond, reduce_dim)
+        
+        self.reduce = nn.Linear(768, reduce_dim)
+
+        self.prompt_list = get_prompt_list(prompt)     
+
+        # precomputed prompts
+        import pickle
+        if isfile('precomputed_prompt_vectors.pickle'):
+            precomp = pickle.load(open('precomputed_prompt_vectors.pickle', 'rb'))
+            self.precomputed_prompts = {k: torch.from_numpy(v) for k, v in precomp.items()}        
+        else:
+            self.precomputed_prompts = dict()
+    
+    def rescaled_pos_emb(self, new_size):
+        assert len(new_size) == 2
+
+        a = self.model.positional_embedding[1:].T.view(1, 768, *self.token_shape)
+        b = nnf.interpolate(a, new_size, mode='bicubic', align_corners=False).squeeze(0).view(768, new_size[0]*new_size[1]).T
+        return torch.cat([self.model.positional_embedding[:1], b])
+
+    def visual_forward(self, x_inp, extract_layers=(), skip=False, mask=None):
+        
+
+        with torch.no_grad():
+
+            inp_size = x_inp.shape[2:]
+
+            if self.n_tokens is not None:
+                stride2 = x_inp.shape[2] // self.n_tokens
+                conv_weight2 = nnf.interpolate(self.model.conv1.weight, (stride2, stride2), mode='bilinear', align_corners=True)
+                x = nnf.conv2d(x_inp, conv_weight2, bias=self.model.conv1.bias, stride=stride2, dilation=self.model.conv1.dilation)
+            else:
+                x = self.model.conv1(x_inp)  # shape = [*, width, grid, grid]
+
+            x = x.reshape(x.shape[0], x.shape[1], -1)  # shape = [*, width, grid ** 2]
+            x = x.permute(0, 2, 1)  # shape = [*, grid ** 2, width]
+
+            x = torch.cat([self.model.class_embedding.to(x.dtype) + torch.zeros(x.shape[0], 1, x.shape[-1], dtype=x.dtype, device=x.device), x], dim=1)  # shape = [*, grid ** 2 + 1, width]
+
+            standard_n_tokens = 50 if self.model.conv1.kernel_size[0] == 32 else 197
+
+            if x.shape[1] != standard_n_tokens:
+                new_shape = int(math.sqrt(x.shape[1]-1))
+                x = x + self.rescaled_pos_emb((new_shape, new_shape)).to(x.dtype)[None,:,:]
+            else:
+                x = x + self.model.positional_embedding.to(x.dtype)
+
+            x = self.model.ln_pre(x)
+
+            x = x.permute(1, 0, 2)  # NLD -> LND
+
+            activations, affinities = [], []
+            for i, res_block in enumerate(self.model.transformer.resblocks):
+                
+                if mask is not None:
+                    mask_layer, mask_type, mask_tensor = mask
+                    if mask_layer == i or mask_layer == 'all':
+                        # import ipdb; ipdb.set_trace()
+                        size = int(math.sqrt(x.shape[0] - 1))
+                        
+                        attn_mask = (mask_type, nnf.interpolate(mask_tensor.unsqueeze(1).float(), (size, size)).view(mask_tensor.shape[0], size * size))
+                        
+                    else:
+                        attn_mask = None
+                else:
+                    attn_mask = None
+
+                x, aff_per_head = forward_multihead_attention(x, res_block, with_aff=True, attn_mask=attn_mask)
+
+                if i in extract_layers:
+                    affinities += [aff_per_head]
+
+                    #if self.n_tokens is not None:
+                    #    activations += [nnf.interpolate(x, inp_size, mode='bilinear', align_corners=True)]
+                    #else:
+                    activations += [x]
+
+                if len(extract_layers) > 0 and i == max(extract_layers) and skip:
+                    print('early skip')
+                    break
+                
+            x = x.permute(1, 0, 2)  # LND -> NLD
+            x = self.model.ln_post(x[:, 0, :])
+
+            if self.model.proj is not None:
+                x = x @ self.model.proj
+
+            return x, activations, affinities
+
+    def sample_prompts(self, words, prompt_list=None):
+
+        prompt_list = prompt_list if prompt_list is not None else self.prompt_list
+
+        prompt_indices = torch.multinomial(torch.ones(len(prompt_list)), len(words), replacement=True)
+        prompts = [prompt_list[i] for i in prompt_indices]
+        return [promt.format(w) for promt, w in zip(prompts, words)]
+
+    def get_cond_vec(self, conditional, batch_size):
+        # compute conditional from a single string
+        if conditional is not None and type(conditional) == str:
+            cond = self.compute_conditional(conditional)
+            cond = cond.repeat(batch_size, 1)
+
+        # compute conditional from string list/tuple
+        elif conditional is not None and type(conditional) in {list, tuple} and type(conditional[0]) == str:
+            assert len(conditional) == batch_size
+            cond = self.compute_conditional(conditional)
+
+        # use conditional directly
+        elif conditional is not None and type(conditional) == torch.Tensor and conditional.ndim == 2:
+            cond = conditional
+
+        # compute conditional from image
+        elif conditional is not None and type(conditional) == torch.Tensor:
+            with torch.no_grad():
+                cond, _, _ = self.visual_forward(conditional)
+        else:
+            raise ValueError('invalid conditional')
+        return cond   
+
+    def compute_conditional(self, conditional):
+        import clip
+
+        dev = next(self.parameters()).device
+
+        if type(conditional) in {list, tuple}:
+            text_tokens = clip.tokenize(conditional).to(dev)
+            cond = self.clip_model.encode_text(text_tokens)
+        else:
+            if conditional in self.precomputed_prompts:
+                cond = self.precomputed_prompts[conditional].float().to(dev)
+            else:
+                text_tokens = clip.tokenize([conditional]).to(dev)
+                cond = self.clip_model.encode_text(text_tokens)[0]
+        
+        if self.shift_vector is not None:
+            return cond + self.shift_vector
+        else:
+            return cond
+
+
+def clip_load_untrained(version):
+    assert version == 'ViT-B/16'
+    from clip.model import CLIP
+    from clip.clip import _MODELS, _download
+    model = torch.jit.load(_download(_MODELS['ViT-B/16'])).eval()
+    state_dict = model.state_dict()
+
+    vision_width = state_dict["visual.conv1.weight"].shape[0]
+    vision_layers = len([k for k in state_dict.keys() if k.startswith("visual.") and k.endswith(".attn.in_proj_weight")])
+    vision_patch_size = state_dict["visual.conv1.weight"].shape[-1]
+    grid_size = round((state_dict["visual.positional_embedding"].shape[0] - 1) ** 0.5)
+    image_resolution = vision_patch_size * grid_size
+    embed_dim = state_dict["text_projection"].shape[1]
+    context_length = state_dict["positional_embedding"].shape[0]
+    vocab_size = state_dict["token_embedding.weight"].shape[0]
+    transformer_width = state_dict["ln_final.weight"].shape[0]
+    transformer_heads = transformer_width // 64
+    transformer_layers = len(set(k.split(".")[2] for k in state_dict if k.startswith(f"transformer.resblocks")))
+
+    return CLIP(embed_dim, image_resolution, vision_layers, vision_width, vision_patch_size, 
+        context_length, vocab_size, transformer_width, transformer_heads, transformer_layers)    
+
+
+class CLIPDensePredT(CLIPDenseBase):
+
+    def __init__(self, version='ViT-B/32', extract_layers=(3, 6, 9), cond_layer=0, reduce_dim=128, n_heads=4, prompt='fixed', 
+                 extra_blocks=0, reduce_cond=None, fix_shift=False,
+                 learn_trans_conv_only=False,  limit_to_clip_only=False, upsample=False, 
+                 add_calibration=False, rev_activations=False, trans_conv=None, n_tokens=None, complex_trans_conv=False):
+        
+        super().__init__(version, reduce_cond, reduce_dim, prompt, n_tokens)
+        # device = 'cpu'
+
+        self.extract_layers = extract_layers
+        self.cond_layer = cond_layer
+        self.limit_to_clip_only = limit_to_clip_only
+        self.process_cond = None
+        self.rev_activations = rev_activations
+        
+        depth = len(extract_layers)
+
+        if add_calibration:
+            self.calibration_conds = 1
+
+        self.upsample_proj = nn.Conv2d(reduce_dim, 1, kernel_size=1) if upsample else None
+
+        self.add_activation1 = True
+
+        self.version = version
+        
+        self.token_shape = {'ViT-B/32': (7, 7), 'ViT-B/16': (14, 14)}[version]
+
+        if fix_shift:
+            # self.shift_vector = nn.Parameter(torch.load(join(dirname(basename(__file__)), 'clip_text_shift_vector.pth')), requires_grad=False)
+            self.shift_vector = nn.Parameter(torch.load(join(dirname(basename(__file__)), 'shift_text_to_vis.pth')), requires_grad=False)
+            # self.shift_vector = nn.Parameter(-1*torch.load(join(dirname(basename(__file__)), 'shift2.pth')), requires_grad=False)
+        else:
+            self.shift_vector = None
+
+        if trans_conv is None:
+            trans_conv_ks = {'ViT-B/32': (32, 32), 'ViT-B/16': (16, 16)}[version]
+        else:
+            # explicitly define transposed conv kernel size
+            trans_conv_ks = (trans_conv, trans_conv)
+
+        if not complex_trans_conv:
+            self.trans_conv = nn.ConvTranspose2d(reduce_dim, 1, trans_conv_ks, stride=trans_conv_ks)
+        else:
+            assert trans_conv_ks[0] == trans_conv_ks[1]
+
+            tp_kernels = (trans_conv_ks[0] // 4, trans_conv_ks[0] // 4)
+
+            self.trans_conv = nn.Sequential(
+                nn.Conv2d(reduce_dim, reduce_dim, kernel_size=3, padding=1),
+                nn.ReLU(),
+                nn.ConvTranspose2d(reduce_dim, reduce_dim // 2, kernel_size=tp_kernels[0], stride=tp_kernels[0]),
+                nn.ReLU(),
+                nn.ConvTranspose2d(reduce_dim // 2, 1, kernel_size=tp_kernels[1], stride=tp_kernels[1]),               
+            )
+
+#        self.trans_conv = nn.ConvTranspose2d(reduce_dim, 1, trans_conv_ks, stride=trans_conv_ks)
+        
+        assert len(self.extract_layers) == depth
+
+        self.reduces = nn.ModuleList([nn.Linear(768, reduce_dim) for _ in range(depth)])
+        self.blocks = nn.ModuleList([nn.TransformerEncoderLayer(d_model=reduce_dim, nhead=n_heads) for _ in range(len(self.extract_layers))])
+        self.extra_blocks = nn.ModuleList([nn.TransformerEncoderLayer(d_model=reduce_dim, nhead=n_heads) for _ in range(extra_blocks)])
+        
+        # refinement and trans conv
+
+        if learn_trans_conv_only:
+            for p in self.parameters():
+                p.requires_grad_(False)
+            
+            for p in self.trans_conv.parameters():
+                p.requires_grad_(True)
+
+        self.prompt_list = get_prompt_list(prompt)
+
+
+    def forward(self, inp_image, conditional=None, return_features=False, mask=None):
+
+        assert type(return_features) == bool
+
+        inp_image = inp_image.to(self.model.positional_embedding.device)
+
+        if mask is not None:
+            raise ValueError('mask not supported')
+
+        # x_inp = normalize(inp_image)
+        x_inp = inp_image
+
+        bs, dev = inp_image.shape[0], x_inp.device
+
+        cond = self.get_cond_vec(conditional, bs)
+
+        visual_q, activations, _ = self.visual_forward(x_inp, extract_layers=[0] + list(self.extract_layers))
+
+        activation1 = activations[0]
+        activations = activations[1:]
+
+        _activations = activations[::-1] if not self.rev_activations else activations
+
+        a = None
+        for i, (activation, block, reduce) in enumerate(zip(_activations, self.blocks, self.reduces)):
+            
+            if a is not None:
+                a = reduce(activation) + a
+            else:
+                a = reduce(activation)
+
+            if i == self.cond_layer:
+                if self.reduce_cond is not None:
+                    cond = self.reduce_cond(cond)
+                
+                a = self.film_mul(cond) * a + self.film_add(cond)
+
+            a = block(a)
+
+        for block in self.extra_blocks:
+            a = a + block(a)
+
+        a = a[1:].permute(1, 2, 0) # rm cls token and -> BS, Feats, Tokens
+
+        size = int(math.sqrt(a.shape[2]))
+
+        a = a.view(bs, a.shape[1], size, size)
+
+        a = self.trans_conv(a)
+
+        if self.n_tokens is not None:
+            a = nnf.interpolate(a, x_inp.shape[2:], mode='bilinear', align_corners=True) 
+
+        if self.upsample_proj is not None:
+            a = self.upsample_proj(a)
+            a = nnf.interpolate(a, x_inp.shape[2:], mode='bilinear')
+
+        if return_features:
+            return a, visual_q, cond, [activation1] + activations
+        else:
+            return a,
+
+
+
+class CLIPDensePredTMasked(CLIPDensePredT):
+
+    def __init__(self, version='ViT-B/32', extract_layers=(3, 6, 9), cond_layer=0, reduce_dim=128, n_heads=4, 
+                 prompt='fixed', extra_blocks=0, reduce_cond=None, fix_shift=False, learn_trans_conv_only=False, 
+                 refine=None, limit_to_clip_only=False, upsample=False, add_calibration=False, n_tokens=None):
+
+        super().__init__(version=version, extract_layers=extract_layers, cond_layer=cond_layer, reduce_dim=reduce_dim, 
+                         n_heads=n_heads, prompt=prompt, extra_blocks=extra_blocks, reduce_cond=reduce_cond, 
+                         fix_shift=fix_shift, learn_trans_conv_only=learn_trans_conv_only,
+                         limit_to_clip_only=limit_to_clip_only, upsample=upsample, add_calibration=add_calibration,
+                         n_tokens=n_tokens)
+
+    def visual_forward_masked(self, img_s, seg_s):
+        return super().visual_forward(img_s, mask=('all', 'cls_token', seg_s))
+
+    def forward(self, img_q, cond_or_img_s, seg_s=None, return_features=False):
+
+        if seg_s is None:
+            cond = cond_or_img_s
+        else:
+            img_s = cond_or_img_s
+
+            with torch.no_grad():
+                cond, _, _ = self.visual_forward_masked(img_s, seg_s)
+
+        return super().forward(img_q, cond, return_features=return_features)
+
+
+
+class CLIPDenseBaseline(CLIPDenseBase):
+
+    def __init__(self, version='ViT-B/32', cond_layer=0, 
+                extract_layer=9, reduce_dim=128, reduce2_dim=None, prompt='fixed', 
+                 reduce_cond=None, limit_to_clip_only=False, n_tokens=None):
+        
+        super().__init__(version, reduce_cond, reduce_dim, prompt, n_tokens)
+        device = 'cpu'
+
+        # self.cond_layer = cond_layer
+        self.extract_layer = extract_layer
+        self.limit_to_clip_only = limit_to_clip_only
+        self.shift_vector = None
+
+        self.token_shape = {'ViT-B/32': (7, 7), 'ViT-B/16': (14, 14)}[version]
+        
+        assert reduce2_dim is not None
+
+        self.reduce2 = nn.Sequential(
+            nn.Linear(reduce_dim, reduce2_dim),
+            nn.ReLU(),
+            nn.Linear(reduce2_dim, reduce_dim)
+        )
+        
+        trans_conv_ks = {'ViT-B/32': (32, 32), 'ViT-B/16': (16, 16)}[version]
+        self.trans_conv = nn.ConvTranspose2d(reduce_dim, 1, trans_conv_ks, stride=trans_conv_ks)
+
+
+    def forward(self, inp_image, conditional=None, return_features=False):
+
+        inp_image = inp_image.to(self.model.positional_embedding.device)
+
+        # x_inp = normalize(inp_image)
+        x_inp = inp_image
+
+        bs, dev = inp_image.shape[0], x_inp.device
+
+        cond = self.get_cond_vec(conditional, bs)
+
+        visual_q, activations, affinities = self.visual_forward(x_inp, extract_layers=[self.extract_layer])
+
+        a = activations[0]
+        a = self.reduce(a)
+        a = self.film_mul(cond) * a + self.film_add(cond)
+
+        if self.reduce2 is not None:
+            a = self.reduce2(a)
+
+        # the original model would execute a transformer block here
+
+        a = a[1:].permute(1, 2, 0) # rm cls token and -> BS, Feats, Tokens
+
+        size = int(math.sqrt(a.shape[2]))
+
+        a = a.view(bs, a.shape[1], size, size)
+        a = self.trans_conv(a)
+
+        if return_features:
+            return a, visual_q, cond, activations
+        else:
+            return a,
+
+
+class CLIPSegMultiLabel(nn.Module):
+
+    def __init__(self, model) -> None:
+        super().__init__()
+
+        from third_party.JoEm.data_loader import get_seen_idx, get_unseen_idx, VOC
+
+        self.pascal_classes = VOC
+
+        from clip.clipseg import CLIPDensePredT
+        from general_utils import load_model
+        # self.clipseg = load_model('rd64-vit16-neg0.2-phrasecut', strict=False)
+        self.clipseg = load_model(model, strict=False)
+        
+        self.clipseg.eval()
+
+    def forward(self, x):
+
+        bs = x.shape[0]
+        out = torch.ones(21, bs, 352, 352).to(x.device) * -10
+
+        for class_id, class_name in enumerate(self.pascal_classes):
+        
+            fac = 3 if class_name == 'background' else 1
+
+            with torch.no_grad():
+                pred = torch.sigmoid(self.clipseg(x, class_name)[0][:,0]) * fac
+
+            out[class_id] += pred
+
+
+        out = out.permute(1, 0, 2, 3)
+
+        return out
+
+        # construct output tensor
+                    

clip/model.py

@@ -0,0 +1,436 @@
+from collections import OrderedDict
+from typing import Tuple, Union
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+
+class Bottleneck(nn.Module):
+    expansion = 4
+
+    def __init__(self, inplanes, planes, stride=1):
+        super().__init__()
+
+        # all conv layers have stride 1. an avgpool is performed after the second convolution when stride > 1
+        self.conv1 = nn.Conv2d(inplanes, planes, 1, bias=False)
+        self.bn1 = nn.BatchNorm2d(planes)
+        self.relu1 = nn.ReLU(inplace=True)
+
+        self.conv2 = nn.Conv2d(planes, planes, 3, padding=1, bias=False)
+        self.bn2 = nn.BatchNorm2d(planes)
+        self.relu2 = nn.ReLU(inplace=True)
+
+        self.avgpool = nn.AvgPool2d(stride) if stride > 1 else nn.Identity()
+
+        self.conv3 = nn.Conv2d(planes, planes * self.expansion, 1, bias=False)
+        self.bn3 = nn.BatchNorm2d(planes * self.expansion)
+        self.relu3 = nn.ReLU(inplace=True)
+
+        self.downsample = None
+        self.stride = stride
+
+        if stride > 1 or inplanes != planes * Bottleneck.expansion:
+            # downsampling layer is prepended with an avgpool, and the subsequent convolution has stride 1
+            self.downsample = nn.Sequential(OrderedDict([
+                ("-1", nn.AvgPool2d(stride)),
+                ("0", nn.Conv2d(inplanes, planes * self.expansion, 1, stride=1, bias=False)),
+                ("1", nn.BatchNorm2d(planes * self.expansion))
+            ]))
+
+    def forward(self, x: torch.Tensor):
+        identity = x
+
+        out = self.relu1(self.bn1(self.conv1(x)))
+        out = self.relu2(self.bn2(self.conv2(out)))
+        out = self.avgpool(out)
+        out = self.bn3(self.conv3(out))
+
+        if self.downsample is not None:
+            identity = self.downsample(x)
+
+        out += identity
+        out = self.relu3(out)
+        return out
+
+
+class AttentionPool2d(nn.Module):
+    def __init__(self, spacial_dim: int, embed_dim: int, num_heads: int, output_dim: int = None):
+        super().__init__()
+        self.positional_embedding = nn.Parameter(torch.randn(spacial_dim ** 2 + 1, embed_dim) / embed_dim ** 0.5)
+        self.k_proj = nn.Linear(embed_dim, embed_dim)
+        self.q_proj = nn.Linear(embed_dim, embed_dim)
+        self.v_proj = nn.Linear(embed_dim, embed_dim)
+        self.c_proj = nn.Linear(embed_dim, output_dim or embed_dim)
+        self.num_heads = num_heads
+
+    def forward(self, x):
+        x = x.flatten(start_dim=2).permute(2, 0, 1)  # NCHW -> (HW)NC
+        x = torch.cat([x.mean(dim=0, keepdim=True), x], dim=0)  # (HW+1)NC
+        x = x + self.positional_embedding[:, None, :].to(x.dtype)  # (HW+1)NC
+        x, _ = F.multi_head_attention_forward(
+            query=x[:1], key=x, value=x,
+            embed_dim_to_check=x.shape[-1],
+            num_heads=self.num_heads,
+            q_proj_weight=self.q_proj.weight,
+            k_proj_weight=self.k_proj.weight,
+            v_proj_weight=self.v_proj.weight,
+            in_proj_weight=None,
+            in_proj_bias=torch.cat([self.q_proj.bias, self.k_proj.bias, self.v_proj.bias]),
+            bias_k=None,
+            bias_v=None,
+            add_zero_attn=False,
+            dropout_p=0,
+            out_proj_weight=self.c_proj.weight,
+            out_proj_bias=self.c_proj.bias,
+            use_separate_proj_weight=True,
+            training=self.training,
+            need_weights=False
+        )
+        return x.squeeze(0)
+
+
+class ModifiedResNet(nn.Module):
+    """
+    A ResNet class that is similar to torchvision's but contains the following changes:
+    - There are now 3 "stem" convolutions as opposed to 1, with an average pool instead of a max pool.
+    - Performs anti-aliasing strided convolutions, where an avgpool is prepended to convolutions with stride > 1
+    - The final pooling layer is a QKV attention instead of an average pool
+    """
+
+    def __init__(self, layers, output_dim, heads, input_resolution=224, width=64):
+        super().__init__()
+        self.output_dim = output_dim
+        self.input_resolution = input_resolution
+
+        # the 3-layer stem
+        self.conv1 = nn.Conv2d(3, width // 2, kernel_size=3, stride=2, padding=1, bias=False)
+        self.bn1 = nn.BatchNorm2d(width // 2)
+        self.relu1 = nn.ReLU(inplace=True)
+        self.conv2 = nn.Conv2d(width // 2, width // 2, kernel_size=3, padding=1, bias=False)
+        self.bn2 = nn.BatchNorm2d(width // 2)
+        self.relu2 = nn.ReLU(inplace=True)
+        self.conv3 = nn.Conv2d(width // 2, width, kernel_size=3, padding=1, bias=False)
+        self.bn3 = nn.BatchNorm2d(width)
+        self.relu3 = nn.ReLU(inplace=True)
+        self.avgpool = nn.AvgPool2d(2)
+
+        # residual layers
+        self._inplanes = width  # this is a *mutable* variable used during construction
+        self.layer1 = self._make_layer(width, layers[0])
+        self.layer2 = self._make_layer(width * 2, layers[1], stride=2)
+        self.layer3 = self._make_layer(width * 4, layers[2], stride=2)
+        self.layer4 = self._make_layer(width * 8, layers[3], stride=2)
+
+        embed_dim = width * 32  # the ResNet feature dimension
+        self.attnpool = AttentionPool2d(input_resolution // 32, embed_dim, heads, output_dim)
+
+    def _make_layer(self, planes, blocks, stride=1):
+        layers = [Bottleneck(self._inplanes, planes, stride)]
+
+        self._inplanes = planes * Bottleneck.expansion
+        for _ in range(1, blocks):
+            layers.append(Bottleneck(self._inplanes, planes))
+
+        return nn.Sequential(*layers)
+
+    def forward(self, x):
+        def stem(x):
+            x = self.relu1(self.bn1(self.conv1(x)))
+            x = self.relu2(self.bn2(self.conv2(x)))
+            x = self.relu3(self.bn3(self.conv3(x)))
+            x = self.avgpool(x)
+            return x
+
+        x = x.type(self.conv1.weight.dtype)
+        x = stem(x)
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+        x = self.layer4(x)
+        x = self.attnpool(x)
+
+        return x
+
+
+class LayerNorm(nn.LayerNorm):
+    """Subclass torch's LayerNorm to handle fp16."""
+
+    def forward(self, x: torch.Tensor):
+        orig_type = x.dtype
+        ret = super().forward(x.type(torch.float32))
+        return ret.type(orig_type)
+
+
+class QuickGELU(nn.Module):
+    def forward(self, x: torch.Tensor):
+        return x * torch.sigmoid(1.702 * x)
+
+
+class ResidualAttentionBlock(nn.Module):
+    def __init__(self, d_model: int, n_head: int, attn_mask: torch.Tensor = None):
+        super().__init__()
+
+        self.attn = nn.MultiheadAttention(d_model, n_head)
+        self.ln_1 = LayerNorm(d_model)
+        self.mlp = nn.Sequential(OrderedDict([
+            ("c_fc", nn.Linear(d_model, d_model * 4)),
+            ("gelu", QuickGELU()),
+            ("c_proj", nn.Linear(d_model * 4, d_model))
+        ]))
+        self.ln_2 = LayerNorm(d_model)
+        self.attn_mask = attn_mask
+
+    def attention(self, x: torch.Tensor):
+        self.attn_mask = self.attn_mask.to(dtype=x.dtype, device=x.device) if self.attn_mask is not None else None
+        return self.attn(x, x, x, need_weights=False, attn_mask=self.attn_mask)[0]
+
+    def forward(self, x: torch.Tensor):
+        x = x + self.attention(self.ln_1(x))
+        x = x + self.mlp(self.ln_2(x))
+        return x
+
+
+class Transformer(nn.Module):
+    def __init__(self, width: int, layers: int, heads: int, attn_mask: torch.Tensor = None):
+        super().__init__()
+        self.width = width
+        self.layers = layers
+        self.resblocks = nn.Sequential(*[ResidualAttentionBlock(width, heads, attn_mask) for _ in range(layers)])
+
+    def forward(self, x: torch.Tensor):
+        return self.resblocks(x)
+
+
+class VisionTransformer(nn.Module):
+    def __init__(self, input_resolution: int, patch_size: int, width: int, layers: int, heads: int, output_dim: int):
+        super().__init__()
+        self.input_resolution = input_resolution
+        self.output_dim = output_dim
+        self.conv1 = nn.Conv2d(in_channels=3, out_channels=width, kernel_size=patch_size, stride=patch_size, bias=False)
+
+        scale = width ** -0.5
+        self.class_embedding = nn.Parameter(scale * torch.randn(width))
+        self.positional_embedding = nn.Parameter(scale * torch.randn((input_resolution // patch_size) ** 2 + 1, width))
+        self.ln_pre = LayerNorm(width)
+
+        self.transformer = Transformer(width, layers, heads)
+
+        self.ln_post = LayerNorm(width)
+        self.proj = nn.Parameter(scale * torch.randn(width, output_dim))
+
+    def forward(self, x: torch.Tensor):
+        x = self.conv1(x)  # shape = [*, width, grid, grid]
+        x = x.reshape(x.shape[0], x.shape[1], -1)  # shape = [*, width, grid ** 2]
+        x = x.permute(0, 2, 1)  # shape = [*, grid ** 2, width]
+        x = torch.cat([self.class_embedding.to(x.dtype) + torch.zeros(x.shape[0], 1, x.shape[-1], dtype=x.dtype, device=x.device), x], dim=1)  # shape = [*, grid ** 2 + 1, width]
+        x = x + self.positional_embedding.to(x.dtype)
+        x = self.ln_pre(x)
+
+        x = x.permute(1, 0, 2)  # NLD -> LND
+        x = self.transformer(x)
+        x = x.permute(1, 0, 2)  # LND -> NLD
+
+        x = self.ln_post(x[:, 0, :])
+
+        if self.proj is not None:
+            x = x @ self.proj
+
+        return x
+
+
+class CLIP(nn.Module):
+    def __init__(self,
+                 embed_dim: int,
+                 # vision
+                 image_resolution: int,
+                 vision_layers: Union[Tuple[int, int, int, int], int],
+                 vision_width: int,
+                 vision_patch_size: int,
+                 # text
+                 context_length: int,
+                 vocab_size: int,
+                 transformer_width: int,
+                 transformer_heads: int,
+                 transformer_layers: int
+                 ):
+        super().__init__()
+
+        self.context_length = context_length
+
+        if isinstance(vision_layers, (tuple, list)):
+            vision_heads = vision_width * 32 // 64
+            self.visual = ModifiedResNet(
+                layers=vision_layers,
+                output_dim=embed_dim,
+                heads=vision_heads,
+                input_resolution=image_resolution,
+                width=vision_width
+            )
+        else:
+            vision_heads = vision_width // 64
+            self.visual = VisionTransformer(
+                input_resolution=image_resolution,
+                patch_size=vision_patch_size,
+                width=vision_width,
+                layers=vision_layers,
+                heads=vision_heads,
+                output_dim=embed_dim
+            )
+
+        self.transformer = Transformer(
+            width=transformer_width,
+            layers=transformer_layers,
+            heads=transformer_heads,
+            attn_mask=self.build_attention_mask()
+        )
+
+        self.vocab_size = vocab_size
+        self.token_embedding = nn.Embedding(vocab_size, transformer_width)
+        self.positional_embedding = nn.Parameter(torch.empty(self.context_length, transformer_width))
+        self.ln_final = LayerNorm(transformer_width)
+
+        self.text_projection = nn.Parameter(torch.empty(transformer_width, embed_dim))
+        self.logit_scale = nn.Parameter(torch.ones([]) * np.log(1 / 0.07))
+
+        self.initialize_parameters()
+
+    def initialize_parameters(self):
+        nn.init.normal_(self.token_embedding.weight, std=0.02)
+        nn.init.normal_(self.positional_embedding, std=0.01)
+
+        if isinstance(self.visual, ModifiedResNet):
+            if self.visual.attnpool is not None:
+                std = self.visual.attnpool.c_proj.in_features ** -0.5
+                nn.init.normal_(self.visual.attnpool.q_proj.weight, std=std)
+                nn.init.normal_(self.visual.attnpool.k_proj.weight, std=std)
+                nn.init.normal_(self.visual.attnpool.v_proj.weight, std=std)
+                nn.init.normal_(self.visual.attnpool.c_proj.weight, std=std)
+
+            for resnet_block in [self.visual.layer1, self.visual.layer2, self.visual.layer3, self.visual.layer4]:
+                for name, param in resnet_block.named_parameters():
+                    if name.endswith("bn3.weight"):
+                        nn.init.zeros_(param)
+
+        proj_std = (self.transformer.width ** -0.5) * ((2 * self.transformer.layers) ** -0.5)
+        attn_std = self.transformer.width ** -0.5
+        fc_std = (2 * self.transformer.width) ** -0.5
+        for block in self.transformer.resblocks:
+            nn.init.normal_(block.attn.in_proj_weight, std=attn_std)
+            nn.init.normal_(block.attn.out_proj.weight, std=proj_std)
+            nn.init.normal_(block.mlp.c_fc.weight, std=fc_std)
+            nn.init.normal_(block.mlp.c_proj.weight, std=proj_std)
+
+        if self.text_projection is not None:
+            nn.init.normal_(self.text_projection, std=self.transformer.width ** -0.5)
+
+    def build_attention_mask(self):
+        # lazily create causal attention mask, with full attention between the vision tokens
+        # pytorch uses additive attention mask; fill with -inf
+        mask = torch.empty(self.context_length, self.context_length)
+        mask.fill_(float("-inf"))
+        mask.triu_(1)  # zero out the lower diagonal
+        return mask
+
+    @property
+    def dtype(self):
+        return self.visual.conv1.weight.dtype
+
+    def encode_image(self, image):
+        return self.visual(image.type(self.dtype))
+
+    def encode_text(self, text):
+        x = self.token_embedding(text).type(self.dtype)  # [batch_size, n_ctx, d_model]
+
+        x = x + self.positional_embedding.type(self.dtype)
+        x = x.permute(1, 0, 2)  # NLD -> LND
+        x = self.transformer(x)
+        x = x.permute(1, 0, 2)  # LND -> NLD
+        x = self.ln_final(x).type(self.dtype)
+
+        # x.shape = [batch_size, n_ctx, transformer.width]
+        # take features from the eot embedding (eot_token is the highest number in each sequence)
+        x = x[torch.arange(x.shape[0]), text.argmax(dim=-1)] @ self.text_projection
+
+        return x
+
+    def forward(self, image, text):
+        image_features = self.encode_image(image)
+        text_features = self.encode_text(text)
+
+        # normalized features
+        image_features = image_features / image_features.norm(dim=1, keepdim=True)
+        text_features = text_features / text_features.norm(dim=1, keepdim=True)
+
+        # cosine similarity as logits
+        logit_scale = self.logit_scale.exp()
+        logits_per_image = logit_scale * image_features @ text_features.t()
+        logits_per_text = logits_per_image.t()
+
+        # shape = [global_batch_size, global_batch_size]
+        return logits_per_image, logits_per_text
+
+
+def convert_weights(model: nn.Module):
+    """Convert applicable model parameters to fp16"""
+
+    def _convert_weights_to_fp16(l):
+        if isinstance(l, (nn.Conv1d, nn.Conv2d, nn.Linear)):
+            l.weight.data = l.weight.data.half()
+            if l.bias is not None:
+                l.bias.data = l.bias.data.half()
+
+        if isinstance(l, nn.MultiheadAttention):
+            for attr in [*[f"{s}_proj_weight" for s in ["in", "q", "k", "v"]], "in_proj_bias", "bias_k", "bias_v"]:
+                tensor = getattr(l, attr)
+                if tensor is not None:
+                    tensor.data = tensor.data.half()
+
+        for name in ["text_projection", "proj"]:
+            if hasattr(l, name):
+                attr = getattr(l, name)
+                if attr is not None:
+                    attr.data = attr.data.half()
+
+    model.apply(_convert_weights_to_fp16)
+
+
+def build_model(state_dict: dict):
+    vit = "visual.proj" in state_dict
+
+    if vit:
+        vision_width = state_dict["visual.conv1.weight"].shape[0]
+        vision_layers = len([k for k in state_dict.keys() if k.startswith("visual.") and k.endswith(".attn.in_proj_weight")])
+        vision_patch_size = state_dict["visual.conv1.weight"].shape[-1]
+        grid_size = round((state_dict["visual.positional_embedding"].shape[0] - 1) ** 0.5)
+        image_resolution = vision_patch_size * grid_size
+    else:
+        counts: list = [len(set(k.split(".")[2] for k in state_dict if k.startswith(f"visual.layer{b}"))) for b in [1, 2, 3, 4]]
+        vision_layers = tuple(counts)
+        vision_width = state_dict["visual.layer1.0.conv1.weight"].shape[0]
+        output_width = round((state_dict["visual.attnpool.positional_embedding"].shape[0] - 1) ** 0.5)
+        vision_patch_size = None
+        assert output_width ** 2 + 1 == state_dict["visual.attnpool.positional_embedding"].shape[0]
+        image_resolution = output_width * 32
+
+    embed_dim = state_dict["text_projection"].shape[1]
+    context_length = state_dict["positional_embedding"].shape[0]
+    vocab_size = state_dict["token_embedding.weight"].shape[0]
+    transformer_width = state_dict["ln_final.weight"].shape[0]
+    transformer_heads = transformer_width // 64
+    transformer_layers = len(set(k.split(".")[2] for k in state_dict if k.startswith("transformer.resblocks")))
+
+    model = CLIP(
+        embed_dim,
+        image_resolution, vision_layers, vision_width, vision_patch_size,
+        context_length, vocab_size, transformer_width, transformer_heads, transformer_layers
+    )
+
+    for key in ["input_resolution", "context_length", "vocab_size"]:
+        if key in state_dict:
+            del state_dict[key]
+
+    convert_weights(model)
+    model.load_state_dict(state_dict)
+    return model.eval()

clip/simple_tokenizer.py

@@ -0,0 +1,132 @@
+import gzip
+import html
+import os
+from functools import lru_cache
+
+import ftfy
+import regex as re
+
+
+@lru_cache()
+def default_bpe():
+    return os.path.join(os.path.dirname(os.path.abspath(__file__)), "bpe_simple_vocab_16e6.txt.gz")
+
+
+@lru_cache()
+def bytes_to_unicode():
+    """
+    Returns list of utf-8 byte and a corresponding list of unicode strings.
+    The reversible bpe codes work on unicode strings.
+    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
+    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
+    This is a signficant percentage of your normal, say, 32K bpe vocab.
+    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
+    And avoids mapping to whitespace/control characters the bpe code barfs on.
+    """
+    bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1))
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8+n)
+            n += 1
+    cs = [chr(n) for n in cs]
+    return dict(zip(bs, cs))
+
+
+def get_pairs(word):
+    """Return set of symbol pairs in a word.
+    Word is represented as tuple of symbols (symbols being variable-length strings).
+    """
+    pairs = set()
+    prev_char = word[0]
+    for char in word[1:]:
+        pairs.add((prev_char, char))
+        prev_char = char
+    return pairs
+
+
+def basic_clean(text):
+    text = ftfy.fix_text(text)
+    text = html.unescape(html.unescape(text))
+    return text.strip()
+
+
+def whitespace_clean(text):
+    text = re.sub(r'\s+', ' ', text)
+    text = text.strip()
+    return text
+
+
+class SimpleTokenizer(object):
+    def __init__(self, bpe_path: str = default_bpe()):
+        self.byte_encoder = bytes_to_unicode()
+        self.byte_decoder = {v: k for k, v in self.byte_encoder.items()}
+        merges = gzip.open(bpe_path).read().decode("utf-8").split('\n')
+        merges = merges[1:49152-256-2+1]
+        merges = [tuple(merge.split()) for merge in merges]
+        vocab = list(bytes_to_unicode().values())
+        vocab = vocab + [v+'</w>' for v in vocab]
+        for merge in merges:
+            vocab.append(''.join(merge))
+        vocab.extend(['<|startoftext|>', '<|endoftext|>'])
+        self.encoder = dict(zip(vocab, range(len(vocab))))
+        self.decoder = {v: k for k, v in self.encoder.items()}
+        self.bpe_ranks = dict(zip(merges, range(len(merges))))
+        self.cache = {'<|startoftext|>': '<|startoftext|>', '<|endoftext|>': '<|endoftext|>'}
+        self.pat = re.compile(r"""<\|startoftext\|>|<\|endoftext\|>|'s|'t|'re|'ve|'m|'ll|'d|[\p{L}]+|[\p{N}]|[^\s\p{L}\p{N}]+""", re.IGNORECASE)
+
+    def bpe(self, token):
+        if token in self.cache:
+            return self.cache[token]
+        word = tuple(token[:-1]) + ( token[-1] + '</w>',)
+        pairs = get_pairs(word)
+
+        if not pairs:
+            return token+'</w>'
+
+        while True:
+            bigram = min(pairs, key = lambda pair: self.bpe_ranks.get(pair, float('inf')))
+            if bigram not in self.bpe_ranks:
+                break
+            first, second = bigram
+            new_word = []
+            i = 0
+            while i < len(word):
+                try:
+                    j = word.index(first, i)
+                    new_word.extend(word[i:j])
+                    i = j
+                except:
+                    new_word.extend(word[i:])
+                    break
+
+                if word[i] == first and i < len(word)-1 and word[i+1] == second:
+                    new_word.append(first+second)
+                    i += 2
+                else:
+                    new_word.append(word[i])
+                    i += 1
+            new_word = tuple(new_word)
+            word = new_word
+            if len(word) == 1:
+                break
+            else:
+                pairs = get_pairs(word)
+        word = ' '.join(word)
+        self.cache[token] = word
+        return word
+
+    def encode(self, text):
+        bpe_tokens = []
+        text = whitespace_clean(basic_clean(text)).lower()
+        for token in re.findall(self.pat, text):
+            token = ''.join(self.byte_encoder[b] for b in token.encode('utf-8'))
+            bpe_tokens.extend(self.encoder[bpe_token] for bpe_token in self.bpe(token).split(' '))
+        return bpe_tokens
+
+    def decode(self, tokens):
+        text = ''.join([self.decoder[token] for token in tokens])
+        text = bytearray([self.byte_decoder[c] for c in text]).decode('utf-8', errors="replace").replace('</w>', ' ')
+        return text

clip/vitseg.py

@@ -0,0 +1,286 @@
+import math
+from posixpath import basename, dirname, join
+# import clip
+from clip.model import convert_weights
+import torch
+import json
+from torch import nn
+from torch.nn import functional as nnf
+from torch.nn.modules import activation
+from torch.nn.modules.activation import ReLU
+from torchvision import transforms
+
+normalize = transforms.Normalize(mean=(0.48145466, 0.4578275, 0.40821073), std=(0.26862954, 0.26130258, 0.27577711))
+
+from torchvision.models import ResNet
+
+
+def process_prompts(conditional, prompt_list, conditional_map):
+    # DEPRECATED
+            
+    # randomly sample a synonym
+    words = [conditional_map[int(i)] for i in conditional]
+    words = [syns[torch.multinomial(torch.ones(len(syns)), 1, replacement=True).item()] for syns in words]
+    words = [w.replace('_', ' ') for w in words]
+
+    if prompt_list is not None:
+        prompt_indices = torch.multinomial(torch.ones(len(prompt_list)), len(words), replacement=True)
+        prompts = [prompt_list[i] for i in prompt_indices]
+    else:
+        prompts = ['a photo of {}'] * (len(words))
+
+    return [promt.format(w) for promt, w in zip(prompts, words)]
+
+
+class VITDenseBase(nn.Module):
+    
+    def rescaled_pos_emb(self, new_size):
+        assert len(new_size) == 2
+
+        a = self.model.positional_embedding[1:].T.view(1, 768, *self.token_shape)
+        b = nnf.interpolate(a, new_size, mode='bicubic', align_corners=False).squeeze(0).view(768, new_size[0]*new_size[1]).T
+        return torch.cat([self.model.positional_embedding[:1], b])
+
+    def visual_forward(self, x_inp, extract_layers=(), skip=False, mask=None):
+        
+        with torch.no_grad():
+
+            x_inp = nnf.interpolate(x_inp, (384, 384))
+
+            x = self.model.patch_embed(x_inp)
+            cls_token = self.model.cls_token.expand(x.shape[0], -1, -1)  # stole cls_tokens impl from Phil Wang, thanks
+            if self.model.dist_token is None:
+                x = torch.cat((cls_token, x), dim=1)
+            else:
+                x = torch.cat((cls_token, self.model.dist_token.expand(x.shape[0], -1, -1), x), dim=1)
+            x = self.model.pos_drop(x + self.model.pos_embed)
+
+            activations = []
+            for i, block in enumerate(self.model.blocks):
+                x = block(x)
+
+                if i in extract_layers:
+                    # permute to be compatible with CLIP
+                    activations += [x.permute(1,0,2)]                
+
+            x = self.model.norm(x)
+            x = self.model.head(self.model.pre_logits(x[:, 0]))
+
+            # again for CLIP compatibility
+            # x = x.permute(1, 0, 2)
+
+        return x, activations, None
+
+    def sample_prompts(self, words, prompt_list=None):
+
+        prompt_list = prompt_list if prompt_list is not None else self.prompt_list
+
+        prompt_indices = torch.multinomial(torch.ones(len(prompt_list)), len(words), replacement=True)
+        prompts = [prompt_list[i] for i in prompt_indices]
+        return [promt.format(w) for promt, w in zip(prompts, words)]
+
+    def get_cond_vec(self, conditional, batch_size):
+        # compute conditional from a single string
+        if conditional is not None and type(conditional) == str:
+            cond = self.compute_conditional(conditional)
+            cond = cond.repeat(batch_size, 1)
+
+        # compute conditional from string list/tuple
+        elif conditional is not None and type(conditional) in {list, tuple} and type(conditional[0]) == str:
+            assert len(conditional) == batch_size
+            cond = self.compute_conditional(conditional)
+
+        # use conditional directly
+        elif conditional is not None and type(conditional) == torch.Tensor and conditional.ndim == 2:
+            cond = conditional
+
+        # compute conditional from image
+        elif conditional is not None and type(conditional) == torch.Tensor:
+            with torch.no_grad():
+                cond, _, _ = self.visual_forward(conditional)
+        else:
+            raise ValueError('invalid conditional')
+        return cond   
+
+    def compute_conditional(self, conditional):
+        import clip
+
+        dev = next(self.parameters()).device
+
+        if type(conditional) in {list, tuple}:
+            text_tokens = clip.tokenize(conditional).to(dev)
+            cond = self.clip_model.encode_text(text_tokens)
+        else:
+            if conditional in self.precomputed_prompts:
+                cond = self.precomputed_prompts[conditional].float().to(dev)
+            else:
+                text_tokens = clip.tokenize([conditional]).to(dev)
+                cond = self.clip_model.encode_text(text_tokens)[0]
+        
+        return cond
+
+
+class VITDensePredT(VITDenseBase):
+
+    def __init__(self, extract_layers=(3, 6, 9), cond_layer=0, reduce_dim=128, n_heads=4, prompt='fixed', 
+                 depth=3, extra_blocks=0, reduce_cond=None, fix_shift=False,
+                 learn_trans_conv_only=False, refine=None, limit_to_clip_only=False, upsample=False, 
+                 add_calibration=False, process_cond=None, not_pretrained=False):
+        super().__init__()
+        # device = 'cpu'
+
+        self.extract_layers = extract_layers
+        self.cond_layer = cond_layer
+        self.limit_to_clip_only = limit_to_clip_only
+        self.process_cond = None
+        
+        if add_calibration:
+            self.calibration_conds = 1
+
+        self.upsample_proj = nn.Conv2d(reduce_dim, 1, kernel_size=1) if upsample else None
+
+        self.add_activation1 = True
+
+        import timm 
+        self.model = timm.create_model('vit_base_patch16_384', pretrained=True)
+        self.model.head = nn.Linear(768, 512 if reduce_cond is None else reduce_cond)
+
+        for p in self.model.parameters():
+            p.requires_grad_(False)
+
+        import clip
+        self.clip_model, _ = clip.load('ViT-B/16', device='cpu', jit=False)
+        # del self.clip_model.visual
+        
+        
+        self.token_shape = (14, 14)
+
+        # conditional
+        if reduce_cond is not None:
+            self.reduce_cond = nn.Linear(512, reduce_cond)
+            for p in self.reduce_cond.parameters():
+                p.requires_grad_(False)
+        else:
+            self.reduce_cond = None
+
+        # self.film = AVAILABLE_BLOCKS['film'](512, 128)
+        self.film_mul = nn.Linear(512 if reduce_cond is None else reduce_cond, reduce_dim)
+        self.film_add = nn.Linear(512 if reduce_cond is None else reduce_cond, reduce_dim)
+        
+        # DEPRECATED
+        # self.conditional_map = {c['id']: c['synonyms'] for c in json.load(open(cond_map))}
+        
+        assert len(self.extract_layers) == depth
+
+        self.reduces = nn.ModuleList([nn.Linear(768, reduce_dim) for _ in range(depth)])
+        self.blocks = nn.ModuleList([nn.TransformerEncoderLayer(d_model=reduce_dim, nhead=n_heads) for _ in range(len(self.extract_layers))])
+        self.extra_blocks = nn.ModuleList([nn.TransformerEncoderLayer(d_model=reduce_dim, nhead=n_heads) for _ in range(extra_blocks)])
+
+        trans_conv_ks = (16, 16)
+        self.trans_conv = nn.ConvTranspose2d(reduce_dim, 1, trans_conv_ks, stride=trans_conv_ks)
+
+        # refinement and trans conv
+
+        if learn_trans_conv_only:
+            for p in self.parameters():
+                p.requires_grad_(False)
+            
+            for p in self.trans_conv.parameters():
+                p.requires_grad_(True)
+
+        if prompt == 'fixed':
+            self.prompt_list = ['a photo of a {}.']
+        elif prompt == 'shuffle':
+            self.prompt_list = ['a photo of a {}.', 'a photograph of a {}.', 'an image of a {}.', '{}.']
+        elif prompt == 'shuffle+':
+            self.prompt_list = ['a photo of a {}.', 'a photograph of a {}.', 'an image of a {}.', '{}.',
+                                'a cropped photo of a {}.', 'a good photo of a {}.', 'a photo of one {}.',
+                                'a bad photo of a {}.', 'a photo of the {}.']
+        elif prompt == 'shuffle_clip':
+            from models.clip_prompts import imagenet_templates
+            self.prompt_list = imagenet_templates
+
+        if process_cond is not None:
+            if process_cond == 'clamp' or process_cond[0] == 'clamp':
+
+                val = process_cond[1] if type(process_cond) in {list, tuple} else 0.2
+
+                def clamp_vec(x):
+                    return torch.clamp(x, -val, val)
+
+                self.process_cond = clamp_vec
+
+            elif process_cond.endswith('.pth'):
+                
+                shift = torch.load(process_cond)
+                def add_shift(x):
+                    return x + shift.to(x.device)
+
+                self.process_cond = add_shift
+
+        import pickle
+        precomp = pickle.load(open('precomputed_prompt_vectors.pickle', 'rb'))
+        self.precomputed_prompts = {k: torch.from_numpy(v) for k, v in precomp.items()}
+
+
+    def forward(self, inp_image, conditional=None, return_features=False, mask=None):
+
+        assert type(return_features) == bool
+
+        # inp_image = inp_image.to(self.model.positional_embedding.device)
+
+        if mask is not None:
+            raise ValueError('mask not supported')
+
+        # x_inp = normalize(inp_image)
+        x_inp = inp_image
+
+        bs, dev = inp_image.shape[0], x_inp.device
+
+        inp_image_size = inp_image.shape[2:]
+
+        cond = self.get_cond_vec(conditional, bs)
+
+        visual_q, activations, _ = self.visual_forward(x_inp, extract_layers=[0] + list(self.extract_layers))
+
+        activation1 = activations[0]
+        activations = activations[1:]
+
+        a = None
+        for i, (activation, block, reduce) in enumerate(zip(activations[::-1], self.blocks, self.reduces)):
+            
+            if a is not None:
+                a = reduce(activation) + a
+            else:
+                a = reduce(activation)
+
+            if i == self.cond_layer:
+                if self.reduce_cond is not None:
+                    cond = self.reduce_cond(cond)
+                
+                a = self.film_mul(cond) * a + self.film_add(cond)
+
+            a = block(a)
+
+        for block in self.extra_blocks:
+            a = a + block(a)
+
+        a = a[1:].permute(1, 2, 0) # rm cls token and -> BS, Feats, Tokens
+
+        size = int(math.sqrt(a.shape[2]))
+
+        a = a.view(bs, a.shape[1], size, size)
+
+        if self.trans_conv is not None:
+            a = self.trans_conv(a)
+
+        if self.upsample_proj is not None:
+            a = self.upsample_proj(a)
+            a = nnf.interpolate(a, x_inp.shape[2:], mode='bilinear')
+
+        a = nnf.interpolate(a, inp_image_size)
+
+        if return_features:
+            return a, visual_q, cond, [activation1] + activations
+        else:
+            return a,

config_colab.yaml

@@ -0,0 +1,15 @@
+clear_output: true
+force_cpu: false
+live_cam_start_active: false
+max_threads: 3
+memory_limit: 0
+output_image_format: png
+output_template: '{file}_{time}'
+output_video_codec: libx264
+output_video_format: mp4
+provider: cuda
+selected_theme: Default
+server_name: ''
+server_port: 0
+server_share: true
+video_quality: 14

installer/installer.py

@@ -0,0 +1,83 @@
+import argparse
+import glob
+import os
+import shutil
+import site
+import subprocess
+import sys
+
+
+script_dir = os.getcwd()
+
+
+def run_cmd(cmd, capture_output=False, env=None):
+    # Run shell commands
+    return subprocess.run(cmd, shell=True, capture_output=capture_output, env=env)
+
+
+def check_env():
+    # If we have access to conda, we are probably in an environment
+    conda_not_exist = run_cmd("conda", capture_output=True).returncode
+    if conda_not_exist:
+        print("Conda is not installed. Exiting...")
+        sys.exit()
+    
+    # Ensure this is a new environment and not the base environment
+    if os.environ["CONDA_DEFAULT_ENV"] == "base":
+        print("Create an environment for this project and activate it. Exiting...")
+        sys.exit()
+
+
+def install_dependencies():
+    # Install Git and clone repo
+    run_cmd("conda install -y -k git")
+    run_cmd("git clone https://github.com/C0untFloyd/roop-unleashed.git")
+    run_cmd("git checkout 8ee085322158c4eeb0cd0126a49949f1acf0f7df")
+    # Install the webui dependencies
+    update_dependencies()
+
+
+def update_dependencies():
+    global MY_PATH
+    
+    os.chdir(MY_PATH)
+	# do a hard reset for to update even if there are local changes
+    run_cmd("git fetch --all")
+    run_cmd("git reset --hard origin/main")
+    run_cmd("git pull")
+    # Installs/Updates dependencies from all requirements.txt
+    run_cmd("python -m pip install -r requirements.txt")
+
+
+def start_app():
+    global MY_PATH
+    
+    os.chdir(MY_PATH)
+    # forward commandline arguments
+    sys.argv.pop(0)
+    args = ' '.join(sys.argv)
+    print("Launching App")
+    run_cmd(f'python run.py {args}')
+
+
+if __name__ == "__main__":
+    global MY_PATH
+    
+    MY_PATH = "roop-unleashed"
+
+    
+    # Verifies we are in a conda environment
+    check_env()
+
+    # If webui has already been installed, skip and run
+    if not os.path.exists(MY_PATH):
+        install_dependencies()
+    else:
+        # moved update from batch to here, because of batch limitations
+        updatechoice = input("Check for Updates? [y/n]").lower()
+        if updatechoice == "y":
+           update_dependencies()
+
+    # Run the model with webui
+    os.chdir(script_dir)
+    start_app()

installer/windows_run.bat

@@ -0,0 +1,80 @@
+@echo off
+REM Please set the following commandline arguments to your prefered settings
+set COMMANDLINE_ARGS=
+
+cd /D "%~dp0"
+
+echo "%CD%"| findstr /C:" " >nul && echo This script relies on Miniconda which can not be silently installed under a path with spaces. && goto end
+
+set PATH=%PATH%;%SystemRoot%\system32
+
+@rem config
+set INSTALL_DIR=%cd%\installer_files
+set CONDA_ROOT_PREFIX=%cd%\installer_files\conda
+set INSTALL_ENV_DIR=%cd%\installer_files\env
+set MINICONDA_DOWNLOAD_URL=https://repo.anaconda.com/miniconda/Miniconda3-latest-Windows-x86_64.exe
+set FFMPEG_DOWNLOAD_URL=https://github.com/GyanD/codexffmpeg/releases/download/2023-06-21-git-1bcb8a7338/ffmpeg-2023-06-21-git-1bcb8a7338-essentials_build.zip
+set INSTALL_FFMPEG_DIR=%cd%\installer_files\ffmpeg
+set conda_exists=F
+
+@rem figure out whether git and conda needs to be installed
+call "%CONDA_ROOT_PREFIX%\_conda.exe" --version >nul 2>&1
+if "%ERRORLEVEL%" EQU "0" set conda_exists=T
+
+@rem (if necessary) install git and conda into a contained environment
+@rem download conda
+if "%conda_exists%" == "F" (
+	echo Downloading Miniconda from %MINICONDA_DOWNLOAD_URL% to %INSTALL_DIR%\miniconda_installer.exe
+
+	mkdir "%INSTALL_DIR%"
+	call curl -Lk "%MINICONDA_DOWNLOAD_URL%" > "%INSTALL_DIR%\miniconda_installer.exe" || ( echo. && echo Miniconda failed to download. && goto end )
+
+	echo Installing Miniconda to %CONDA_ROOT_PREFIX%
+	start /wait "" "%INSTALL_DIR%\miniconda_installer.exe" /InstallationType=JustMe /NoShortcuts=1 /AddToPath=0 /RegisterPython=0 /NoRegistry=1 /S /D=%CONDA_ROOT_PREFIX%
+
+	@rem test the conda binary
+	echo Miniconda version:
+	call "%CONDA_ROOT_PREFIX%\_conda.exe" --version || ( echo. && echo Miniconda not found. && goto end )
+)
+
+@rem create the installer env
+if not exist "%INSTALL_ENV_DIR%" (
+  echo Packages to install: %PACKAGES_TO_INSTALL%
+  call "%CONDA_ROOT_PREFIX%\_conda.exe" create --no-shortcuts -y -k --prefix "%INSTALL_ENV_DIR%" python=3.10 || ( echo. && echo Conda environment creation failed. && goto end )
+)
+
+if not exist "%INSTALL_FFMPEG_DIR%" (
+	echo Downloading ffmpeg from %FFMPEG_DOWNLOAD_URL% to %INSTALL_DIR%
+ 	call curl -Lk "%FFMPEG_DOWNLOAD_URL%" > "%INSTALL_DIR%\ffmpeg.zip" || ( echo. && echo ffmpeg failed to download. && goto end )
+	call powershell -command "Expand-Archive -Force '%INSTALL_DIR%\ffmpeg.zip' '%INSTALL_DIR%\'"
+
+	cd "installer_files"
+	setlocal EnableExtensions EnableDelayedExpansion
+
+	for /f "tokens=*" %%f in ('dir /s /b /ad "ffmpeg*"') do (
+		ren "%%f" "ffmpeg"
+	)
+	endlocal
+	setx PATH "%INSTALL_FFMPEG_DIR%\bin\;%PATH%"
+	echo To use videos, you need to restart roop after this installation. 
+	cd ..
+)
+
+@rem check if conda environment was actually created
+if not exist "%INSTALL_ENV_DIR%\python.exe" ( echo. && echo ERROR: Conda environment is empty. && goto end )
+
+@rem activate installer env
+call "%CONDA_ROOT_PREFIX%\condabin\conda.bat" activate "%INSTALL_ENV_DIR%" || ( echo. && echo Miniconda hook not found. && goto end )
+
+@rem setup installer env
+echo Launching roop unleashed - please edit windows_run.bat to customize commandline arguments
+call python installer.py %COMMANDLINE_ARGS%
+
+echo.
+echo Done!
+
+:end
+pause
+
+
+

mypy.ini

@@ -0,0 +1,7 @@
+[mypy]
+check_untyped_defs = True
+disallow_any_generics = True
+disallow_untyped_calls = True
+disallow_untyped_defs = True
+ignore_missing_imports = True
+strict_optional = False

requirements.txt

@@ -0,0 +1,24 @@
+--extra-index-url https://download.pytorch.org/whl/cu118
+
+numpy==1.24.2
+gradio>=3.41.2
+opencv-python==4.7.0.72
+onnx==1.14.0
+insightface==0.7.3
+psutil==5.9.5
+pillow==9.5.0
+torch==2.0.1+cu118; sys_platform != 'darwin'
+torch==2.0.1; sys_platform == 'darwin'
+torchvision==0.15.2+cu118; sys_platform != 'darwin'
+torchvision==0.15.2; sys_platform == 'darwin'
+onnxruntime==1.15.0; sys_platform == 'darwin' and platform_machine != 'arm64'
+onnxruntime-silicon==1.13.1; sys_platform == 'darwin' and platform_machine == 'arm64'
+onnxruntime-gpu==1.15.0; sys_platform != 'darwin'
+tensorflow==2.13.0
+protobuf==4.23.2
+tqdm==4.65.0
+codeformer-pip==0.0.4
+gfpgan==1.3.8
+ftfy
+regex
+pyvirtualcam

roop-unleashed.ipynb

@@ -0,0 +1,184 @@
+{
+  "nbformat": 4,
+  "nbformat_minor": 0,
+  "metadata": {
+    "colab": {
+      "provenance": [],
+      "gpuType": "T4",
+      "collapsed_sections": [
+        "UdQ1VHdI8lCf"
+      ]
+    },
+    "kernelspec": {
+      "name": "python3",
+      "display_name": "Python 3"
+    },
+    "language_info": {
+      "name": "python"
+    },
+    "accelerator": "GPU"
+  },
+  "cells": [
+    {
+      "cell_type": "markdown",
+      "source": [
+        "# Colab for roop-unleashed - Gradio version\n",
+        "https://github.com/C0untFloyd/roop-unleashed\n"
+      ],
+      "metadata": {
+        "id": "G9BdiCppV6AS"
+      }
+    },
+    {
+      "cell_type": "markdown",
+      "source": [
+        "Installing & preparing requirements"
+      ],
+      "metadata": {
+        "id": "0ZYRNb0AWLLW"
+      }
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "t1yPuhdySqCq"
+      },
+      "outputs": [],
+      "source": [
+        "!git clone https://github.com/C0untFloyd/roop-unleashed.git\n",
+        "%cd roop-unleashed\n",
+        "!mv config_colab.yaml config.yaml\n",
+        "!pip install pip install -r requirements.txt"
+      ]
+    },
+    {
+      "cell_type": "markdown",
+      "source": [
+        "Running roop-unleashed with default config"
+      ],
+      "metadata": {
+        "id": "u_4JQiSlV9Fi"
+      }
+    },
+    {
+      "cell_type": "code",
+      "source": [
+        "!python run.py"
+      ],
+      "metadata": {
+        "id": "Is6U2huqSzLE"
+      },
+      "execution_count": null,
+      "outputs": []
+    },
+    {
+      "cell_type": "markdown",
+      "source": [
+        "### Download generated images folder\n",
+        "(only needed if you want to zip the generated output)"
+      ],
+      "metadata": {
+        "id": "UdQ1VHdI8lCf"
+      }
+    },
+    {
+      "cell_type": "code",
+      "source": [
+        "import shutil\n",
+        "import os\n",
+        "from google.colab import files\n",
+        "\n",
+        "def zip_directory(directory_path, zip_path):\n",
+        "    shutil.make_archive(zip_path, 'zip', directory_path)\n",
+        "\n",
+        "# Set the directory path you want to download\n",
+        "directory_path = '/content/roop-unleashed/output'\n",
+        "\n",
+        "# Set the zip file name\n",
+        "zip_filename = 'fake_output.zip'\n",
+        "\n",
+        "# Zip the directory\n",
+        "zip_directory(directory_path, zip_filename)\n",
+        "\n",
+        "# Download the zip file\n",
+        "files.download(zip_filename+'.zip')\n"
+      ],
+      "metadata": {
+        "colab": {
+          "base_uri": "https://localhost:8080/",
+          "height": 17
+        },
+        "id": "oYjWveAmw10X",
+        "outputId": "5b4c3650-f951-434a-c650-5525a8a70c1e"
+      },
+      "execution_count": null,
+      "outputs": [
+        {
+          "output_type": "display_data",
+          "data": {
+            "text/plain": [
+              "<IPython.core.display.Javascript object>"
+            ],
+            "application/javascript": [
+              "\n",
+              "    async function download(id, filename, size) {\n",
+              "      if (!google.colab.kernel.accessAllowed) {\n",
+              "        return;\n",
+              "      }\n",
+              "      const div = document.createElement('div');\n",
+              "      const label = document.createElement('label');\n",
+              "      label.textContent = `Downloading \"${filename}\": `;\n",
+              "      div.appendChild(label);\n",
+              "      const progress = document.createElement('progress');\n",
+              "      progress.max = size;\n",
+              "      div.appendChild(progress);\n",
+              "      document.body.appendChild(div);\n",
+              "\n",
+              "      const buffers = [];\n",
+              "      let downloaded = 0;\n",
+              "\n",
+              "      const channel = await google.colab.kernel.comms.open(id);\n",
+              "      // Send a message to notify the kernel that we're ready.\n",
+              "      channel.send({})\n",
+              "\n",
+              "      for await (const message of channel.messages) {\n",
+              "        // Send a message to notify the kernel that we're ready.\n",
+              "        channel.send({})\n",
+              "        if (message.buffers) {\n",
+              "          for (const buffer of message.buffers) {\n",
+              "            buffers.push(buffer);\n",
+              "            downloaded += buffer.byteLength;\n",
+              "            progress.value = downloaded;\n",
+              "          }\n",
+              "        }\n",
+              "      }\n",
+              "      const blob = new Blob(buffers, {type: 'application/binary'});\n",
+              "      const a = document.createElement('a');\n",
+              "      a.href = window.URL.createObjectURL(blob);\n",
+              "      a.download = filename;\n",
+              "      div.appendChild(a);\n",
+              "      a.click();\n",
+              "      div.remove();\n",
+              "    }\n",
+              "  "
+            ]
+          },
+          "metadata": {}
+        },
+        {
+          "output_type": "display_data",
+          "data": {
+            "text/plain": [
+              "<IPython.core.display.Javascript object>"
+            ],
+            "application/javascript": [
+              "download(\"download_789eab11-93d2-4880-adf3-6aceee0cc5f9\", \"fake_output.zip.zip\", 80125)"
+            ]
+          },
+          "metadata": {}
+        }
+      ]
+    }
+  ]
+}

roop/ProcessEntry.py

@@ -0,0 +1,7 @@
+class ProcessEntry:
+    def __init__(self, filename: str, start: int, end: int, fps: float):
+        self.filename = filename
+        self.finalname = None
+        self.startframe = start
+        self.endframe = end
+        self.fps = fps

roop/ProcessMgr.py

@@ -0,0 +1,432 @@
+import os
+import cv2 
+import numpy as np
+import psutil
+
+from roop.ProcessOptions import ProcessOptions
+
+from roop.face_util import get_first_face, get_all_faces
+from roop.utilities import compute_cosine_distance, get_device, str_to_class
+
+from typing import Any, List, Callable
+from roop.typing import Frame
+from concurrent.futures import ThreadPoolExecutor, as_completed
+from threading import Thread, Lock
+from queue import Queue
+from tqdm import tqdm
+from roop.ffmpeg_writer import FFMPEG_VideoWriter
+import roop.globals
+
+
+def create_queue(temp_frame_paths: List[str]) -> Queue[str]:
+    queue: Queue[str] = Queue()
+    for frame_path in temp_frame_paths:
+        queue.put(frame_path)
+    return queue
+
+
+def pick_queue(queue: Queue[str], queue_per_future: int) -> List[str]:
+    queues = []
+    for _ in range(queue_per_future):
+        if not queue.empty():
+            queues.append(queue.get())
+    return queues
+
+
+class ProcessMgr():
+    input_face_datas = []
+    target_face_datas = []
+
+    processors = []
+    options : ProcessOptions = None
+    
+    num_threads = 1
+    current_index = 0
+    processing_threads = 1
+    buffer_wait_time = 0.1
+
+    lock = Lock()
+
+    frames_queue = None
+    processed_queue = None
+
+    videowriter= None
+
+    progress_gradio = None
+    total_frames = 0
+
+    
+
+    # 5-point template constant for face alignment - don't ask
+    insight_arcface_dst = np.array(
+            [[38.2946, 51.6963], [73.5318, 51.5014], [56.0252, 71.7366],
+            [41.5493, 92.3655], [70.7299, 92.2041]],
+            dtype=np.float32)  
+
+    plugins =  { 
+    'faceswap'      : 'FaceSwapInsightFace',
+    'mask_clip2seg' : 'Mask_Clip2Seg',
+    'codeformer'    : 'Enhance_CodeFormer',
+    'gfpgan'        : 'Enhance_GFPGAN',
+    'dmdnet'        : 'Enhance_DMDNet',
+    }
+
+    def __init__(self, progress):
+        if progress is not None:
+            self.progress_gradio = progress
+
+
+    def initialize(self, input_faces, target_faces, options):
+        self.input_face_datas = input_faces
+        self.target_face_datas = target_faces
+        self.options = options
+
+        processornames = options.processors.split(",")
+        devicename = get_device()
+        if len(self.processors) < 1:
+            for pn in processornames:
+                classname = self.plugins[pn]
+                module = 'roop.processors.' + classname
+                p = str_to_class(module, classname)
+                p.Initialize(devicename)
+                self.processors.append(p)
+        else:
+            for i in range(len(self.processors) -1, -1, -1):
+                if not self.processors[i].processorname in processornames:
+                    self.processors[i].Release()
+                    del self.processors[i]
+
+            for i,pn in enumerate(processornames):
+                if i >= len(self.processors) or self.processors[i].processorname != pn:
+                    p = None
+                    classname = self.plugins[pn]
+                    module = 'roop.processors.' + classname
+                    p = str_to_class(module, classname)
+                    p.Initialize(devicename)
+                    if p is not None:
+                        self.processors.insert(i, p)
+
+
+
+    def run_batch(self, source_files, target_files, threads:int = 1):
+        progress_bar_format = '{l_bar}{bar}| {n_fmt}/{total_fmt} [{elapsed}<{remaining}, {rate_fmt}{postfix}]'
+        self.total_frames = len(source_files)
+        self.num_threads = threads
+        with tqdm(total=self.total_frames, desc='Processing', unit='frame', dynamic_ncols=True, bar_format=progress_bar_format) as progress:
+            with ThreadPoolExecutor(max_workers=threads) as executor:
+                futures = []
+                queue = create_queue(source_files)
+                queue_per_future = max(len(source_files) // threads, 1)
+                while not queue.empty():
+                    future = executor.submit(self.process_frames, source_files, target_files, pick_queue(queue, queue_per_future), lambda: self.update_progress(progress))
+                    futures.append(future)
+                for future in as_completed(futures):
+                    future.result()
+
+
+    def process_frames(self, source_files: List[str], target_files: List[str], current_files, update: Callable[[], None]) -> None:
+        for f in current_files:
+            if not roop.globals.processing:
+                return
+            
+            temp_frame = cv2.imread(f)
+            if temp_frame is not None:
+                resimg = self.process_frame(temp_frame)
+                if resimg is not None:
+                    i = source_files.index(f)
+                    cv2.imwrite(target_files[i], resimg)
+            if update:
+                update()
+
+
+
+    def read_frames_thread(self, cap, frame_start, frame_end, num_threads):
+        num_frame = 0
+        total_num = frame_end - frame_start
+        if frame_start > 0:
+            cap.set(cv2.CAP_PROP_POS_FRAMES,frame_start)
+
+        while True and roop.globals.processing:
+            ret, frame = cap.read()
+            if not ret:
+                break
+                
+            self.frames_queue[num_frame % num_threads].put(frame, block=True)
+            num_frame += 1
+            if num_frame == total_num:
+                break
+
+        for i in range(num_threads):
+            self.frames_queue[i].put(None)
+
+
+
+    def process_videoframes(self, threadindex, progress) -> None:
+        while True:
+            frame = self.frames_queue[threadindex].get()
+            if frame is None:
+                self.processing_threads -= 1
+                self.processed_queue[threadindex].put(None)
+                return
+            else:
+                resimg = self.process_frame(frame)
+                self.processed_queue[threadindex].put(resimg)
+                del frame
+                progress()
+
+
+    def write_frames_thread(self):
+        nextindex = 0
+        num_producers = self.num_threads
+        
+        while True:
+            frame = self.processed_queue[nextindex % self.num_threads].get()
+            nextindex += 1
+            if frame is not None:
+                self.videowriter.write_frame(frame)
+                del frame
+            else:
+                num_producers -= 1
+                if num_producers < 1:
+                    return
+            
+
+
+    def run_batch_inmem(self, source_video, target_video, frame_start, frame_end, fps, threads:int = 1, skip_audio=False):
+        cap = cv2.VideoCapture(source_video)
+        # frame_count = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
+        frame_count = (frame_end - frame_start) + 1
+        width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
+        height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
+
+        self.total_frames = frame_count
+        self.num_threads = threads
+
+        self.processing_threads = self.num_threads
+        self.frames_queue = []
+        self.processed_queue = []
+        for _ in range(threads):
+            self.frames_queue.append(Queue(1))
+            self.processed_queue.append(Queue(1))
+
+        self.videowriter =  FFMPEG_VideoWriter(target_video, (width, height), fps, codec=roop.globals.video_encoder, crf=roop.globals.video_quality, audiofile=None)
+        if not skip_audio and frame_start > 0:
+            print('Writing offset frames')
+            num_write = frame_start
+            cap.set(cv2.CAP_PROP_POS_FRAMES,frame_start)
+            ret, frame = cap.read()
+            fake_frame = self.process_frame(frame)
+            while num_write > 0:
+                self.videowriter.write_frame(fake_frame)
+                num_write -= 1           
+
+        readthread = Thread(target=self.read_frames_thread, args=(cap, frame_start, frame_end, threads))
+        readthread.start()
+
+        writethread = Thread(target=self.write_frames_thread)
+        writethread.start()
+
+        progress_bar_format = '{l_bar}{bar}| {n_fmt}/{total_fmt} [{elapsed}<{remaining}, {rate_fmt}{postfix}]'
+        with tqdm(total=self.total_frames, desc='Processing', unit='frames', dynamic_ncols=True, bar_format=progress_bar_format) as progress:
+            with ThreadPoolExecutor(thread_name_prefix='swap_proc', max_workers=self.num_threads) as executor:
+                futures = []
+                
+                for threadindex in range(threads):
+                    future = executor.submit(self.process_videoframes, threadindex, lambda: self.update_progress(progress))
+                    futures.append(future)
+                
+                for future in as_completed(futures):
+                    future.result()
+        # wait for the task to complete
+        readthread.join()
+        writethread.join()
+        cap.release()
+        self.videowriter.close()
+        self.frames_queue.clear()
+        self.processed_queue.clear()
+
+
+
+
+    def update_progress(self, progress: Any = None) -> None:
+        process = psutil.Process(os.getpid())
+        memory_usage = process.memory_info().rss / 1024 / 1024 / 1024
+        msg = 'memory_usage: ' + '{:.2f}'.format(memory_usage).zfill(5) + f' GB execution_threads {self.num_threads}'
+        progress.set_postfix({
+            'memory_usage': '{:.2f}'.format(memory_usage).zfill(5) + 'GB',
+            'execution_threads': self.num_threads
+        })
+        progress.update(1)
+        self.progress_gradio((progress.n, self.total_frames), desc='Processing', total=self.total_frames, unit='frames')
+
+
+
+       
+
+    def process_frame(self, frame:Frame):
+        if len(self.input_face_datas) < 1:
+            return frame
+    
+        temp_frame = frame.copy()
+
+        if self.options.swap_mode == "first":
+            face = get_first_face(frame)
+            if face is None:
+                return frame
+            return self.process_face(self.options.selected_index, face, temp_frame)
+
+        else:
+            faces = get_all_faces(frame)
+            if faces is None:
+                return frame
+            
+            if self.options.swap_mode == "all":
+                for face in faces:
+                    temp_frame = self.process_face(self.options.selected_index, face, temp_frame)
+                    del face
+                return temp_frame
+            
+            elif self.options.swap_mode == "selected":
+                for i,tf in enumerate(self.target_face_datas):
+                    for face in faces:
+                        if compute_cosine_distance(tf.embedding, face.embedding) <= self.options.face_distance_threshold:
+                            if i < len(self.input_face_datas):
+                                temp_frame = self.process_face(i, face, temp_frame)
+                            break
+                        del face
+
+            elif self.options.swap_mode == "all_female" or self.options.swap_mode == "all_male":
+                gender = 'F' if self.swap_mode == "all_female" else 'M'
+                for face in faces:
+                    if face.sex == gender:
+                        temp_frame = self.process_face(self.options.selected_index, face, temp_frame)
+                    del face
+
+        return temp_frame
+
+
+    def process_face(self,face_index, target_face, frame:Frame):
+        enhanced_frame = None
+        img_mask = None
+        for p in self.processors:
+            if p.type == 'swap':
+                fake_frame = p.Run(self.input_face_datas[face_index], target_face, frame)
+                scale_factor = 0.0
+            elif p.type == 'mask':
+                start_x, start_y, end_x, end_y = map(int, target_face['bbox'])
+                orig_frame = frame[start_y:end_y, start_x:end_x]
+                img_mask = p.Run(orig_frame, self.options.masking_text)
+            else:
+                enhanced_frame, scale_factor = p.Run(self.input_face_datas[face_index], target_face, fake_frame)
+
+        upscale = 512
+        orig_width = fake_frame.shape[1]
+        fake_frame = cv2.resize(fake_frame, (upscale, upscale), cv2.INTER_CUBIC)
+        mask_top = self.input_face_datas[face_index].mask_top
+        if enhanced_frame is None:
+            scale_factor = int(upscale / orig_width)
+            result = self.paste_upscale(fake_frame, fake_frame, target_face.matrix, frame, scale_factor, mask_top)
+        else:
+            result = self.paste_upscale(fake_frame, enhanced_frame, target_face.matrix, frame, scale_factor, mask_top)
+        if img_mask is not None:
+            target = result[start_y:end_y, start_x:end_x]
+            img_mask = cv2.resize(img_mask, (target.shape[1], target.shape[0]))
+            img_mask = np.reshape(img_mask, [img_mask.shape[0],img_mask.shape[1],1])
+    
+            target = target.astype(np.float32)
+            clip = (1-img_mask) * target
+            clip += img_mask * orig_frame.astype(np.float32)
+            result[start_y:end_y, start_x:end_x] = clip
+            return result
+
+        return result
+
+        
+
+
+    
+    
+    # Paste back adapted from here
+    # https://github.com/fAIseh00d/refacer/blob/main/refacer.py
+    # which is revised insightface paste back code
+
+    def paste_upscale(self, fake_face, upsk_face, M, target_img, scale_factor, mask_top):
+        M_scale = M * scale_factor
+        IM = cv2.invertAffineTransform(M_scale)
+
+        face_matte = np.full((target_img.shape[0],target_img.shape[1]), 255, dtype=np.uint8)
+        ##Generate white square sized as a upsk_face
+        img_matte = np.full((upsk_face.shape[0],upsk_face.shape[1]), 255, dtype=np.uint8)
+        if mask_top > 0:
+            img_matte[:mask_top,:] = 0
+
+        ##Transform white square back to target_img
+        img_matte = cv2.warpAffine(img_matte, IM, (target_img.shape[1], target_img.shape[0]), flags=cv2.INTER_NEAREST, borderValue=0.0) 
+        ##Blacken the edges of face_matte by 1 pixels (so the mask in not expanded on the image edges)
+        img_matte[:1,:] = img_matte[-1:,:] = img_matte[:,:1] = img_matte[:,-1:] = 0
+
+        #Detect the affine transformed white area
+        mask_h_inds, mask_w_inds = np.where(img_matte==255) 
+        #Calculate the size (and diagonal size) of transformed white area width and height boundaries
+        mask_h = np.max(mask_h_inds) - np.min(mask_h_inds) 
+        mask_w = np.max(mask_w_inds) - np.min(mask_w_inds)
+        mask_size = int(np.sqrt(mask_h*mask_w))
+        #Calculate the kernel size for eroding img_matte by kernel (insightface empirical guess for best size was max(mask_size//10,10))
+        # k = max(mask_size//12, 8)
+        k = max(mask_size//10, 10)
+        kernel = np.ones((k,k),np.uint8)
+        img_matte = cv2.erode(img_matte,kernel,iterations = 1)
+        #Calculate the kernel size for blurring img_matte by blur_size (insightface empirical guess for best size was max(mask_size//20, 5))
+        # k = max(mask_size//24, 4) 
+        k = max(mask_size//20, 5) 
+        kernel_size = (k, k)
+        blur_size = tuple(2*i+1 for i in kernel_size)
+        img_matte = cv2.GaussianBlur(img_matte, blur_size, 0)
+        
+        #Normalize images to float values and reshape
+        img_matte = img_matte.astype(np.float32)/255
+        face_matte = face_matte.astype(np.float32)/255
+        img_matte = np.minimum(face_matte, img_matte)
+        img_matte = np.reshape(img_matte, [img_matte.shape[0],img_matte.shape[1],1]) 
+        ##Transform upcaled face back to target_img
+        paste_face = cv2.warpAffine(upsk_face, IM, (target_img.shape[1], target_img.shape[0]), borderMode=cv2.BORDER_REPLICATE)
+        if upsk_face is not fake_face:
+            fake_face = cv2.warpAffine(fake_face, IM, (target_img.shape[1], target_img.shape[0]), borderMode=cv2.BORDER_REPLICATE)
+            paste_face = cv2.addWeighted(paste_face, self.options.blend_ratio, fake_face, 1.0 - self.options.blend_ratio, 0)
+
+        ##Re-assemble image
+        paste_face = img_matte * paste_face
+        paste_face = paste_face + (1-img_matte) * target_img.astype(np.float32)
+        del img_matte
+        del face_matte
+        del upsk_face
+        del fake_face
+        return paste_face.astype(np.uint8)
+
+
+    def process_mask(self, frame:Frame, target:Frame):
+        for p in self.processors:
+            if p.type == 'mask':
+                img_mask = p.Run(frame, self.options.masking_text)
+                img_mask = cv2.resize(img_mask, (target.shape[1], target.shape[0]))
+                img_mask = np.reshape(img_mask, [img_mask.shape[0],img_mask.shape[1],1])
+       
+                target = target.astype(np.float32)
+                result = (1-img_mask) * target
+                result += img_mask * frame.astype(np.float32)
+                return np.uint8(result)
+
+            
+
+
+    def unload_models():
+        pass
+
+
+    def release_resources(self):
+        for p in self.processors:
+            p.Release()
+        self.processors.clear()
+
+        
+

roop/ProcessOptions.py

@@ -0,0 +1,9 @@
+class ProcessOptions:
+
+    def __init__(self,processors, face_distance,  blend_ratio, swap_mode, selected_index, masking_text):
+        self.processors = processors
+        self.face_distance_threshold = face_distance
+        self.blend_ratio = blend_ratio
+        self.swap_mode = swap_mode
+        self.selected_index = selected_index
+        self.masking_text = masking_text

roop/capturer.py

@@ -0,0 +1,30 @@
+from typing import Optional
+import cv2
+
+from roop.typing import Frame
+
+def get_image_frame(filename: str):
+    try:
+        frame = cv2.imread(filename)
+        return frame
+    except:
+        print(f"Exception reading {filename}")
+    return None
+
+    
+def get_video_frame(video_path: str, frame_number: int = 0) -> Optional[Frame]:
+    capture = cv2.VideoCapture(video_path)
+    frame_total = capture.get(cv2.CAP_PROP_FRAME_COUNT)
+    capture.set(cv2.CAP_PROP_POS_FRAMES, min(frame_total, frame_number - 1))
+    has_frame, frame = capture.read()
+    capture.release()
+    if has_frame:
+        return frame
+    return None
+
+
+def get_video_frame_total(video_path: str) -> int:
+    capture = cv2.VideoCapture(video_path)
+    video_frame_total = int(capture.get(cv2.CAP_PROP_FRAME_COUNT))
+    capture.release()
+    return video_frame_total

roop/core.py

@@ -0,0 +1,385 @@
+#!/usr/bin/env python3
+
+import os
+import sys
+import shutil
+# single thread doubles cuda performance - needs to be set before torch import
+if any(arg.startswith('--execution-provider') for arg in sys.argv):
+    os.environ['OMP_NUM_THREADS'] = '1'
+
+import warnings
+from typing import List
+import platform
+import signal
+import argparse
+import torch
+import onnxruntime
+import tensorflow
+import pathlib
+
+from time import time
+
+import roop.globals
+import roop.metadata
+import roop.utilities as util
+import roop.ui as ui
+from settings import Settings
+from roop.face_util import extract_face_images
+from roop.ProcessEntry import ProcessEntry
+from roop.ProcessMgr import ProcessMgr
+from roop.ProcessOptions import ProcessOptions
+
+
+clip_text = None
+
+call_display_ui = None
+
+process_mgr = None
+
+
+if 'ROCMExecutionProvider' in roop.globals.execution_providers:
+    del torch
+
+warnings.filterwarnings('ignore', category=FutureWarning, module='insightface')
+warnings.filterwarnings('ignore', category=UserWarning, module='torchvision')
+
+
+def parse_args() -> None:
+    signal.signal(signal.SIGINT, lambda signal_number, frame: destroy())
+    program = argparse.ArgumentParser(formatter_class=lambda prog: argparse.HelpFormatter(prog, max_help_position=100))
+    program.add_argument('-s', '--source', help='select a source image', dest='source_path')
+    program.add_argument('-t', '--target', help='select a target image or video', dest='target_path')
+    program.add_argument('-o', '--output', help='select output file or directory', dest='output_path')
+    program.add_argument('-f', '--folder', help='select a target folder with images or videos to batch process', dest='target_folder_path')
+    program.add_argument('--frame-processor', help='frame processors (choices: face_swapper, face_enhancer, ...)', dest='frame_processor', default=['face_swapper'], nargs='+')
+    program.add_argument('--keep-fps', help='keep target fps', dest='keep_fps', action='store_true')
+    program.add_argument('--keep-frames', help='keep temporary frames', dest='keep_frames', action='store_true')
+    program.add_argument('--skip-audio', help='skip target audio', dest='skip_audio', action='store_true')
+    program.add_argument('--many-faces', help='process every face', dest='many_faces', action='store_true')
+    program.add_argument('--source-face_index', help='index position of source face in image', dest='source_face_index', type=int, default=0)
+    program.add_argument('--target-face_index', help='index position of target face in image', dest='target_face_index', type=int, default=0)
+    program.add_argument('--video-encoder', help='adjust output video encoder', dest='video_encoder', default='libx264', choices=['libx264', 'libx265', 'libvpx-vp9'])
+    program.add_argument('--video-quality', help='adjust output video quality', dest='video_quality', type=int, default=18, choices=range(52), metavar='[0-51]')
+    program.add_argument('--max-memory', help='maximum amount of RAM in GB', dest='max_memory', type=int, default=suggest_max_memory())
+    program.add_argument('--execution-provider', help='available execution provider (choices: cpu, ...)', dest='execution_provider', default=['cpu'], choices=suggest_execution_providers(), nargs='+')
+    program.add_argument('--execution-threads', help='number of execution threads', dest='execution_threads', type=int, default=suggest_execution_threads())
+    program.add_argument('-v', '--version', action='version', version=f'{roop.metadata.name} {roop.metadata.version}')
+
+    args = program.parse_args()
+
+    roop.globals.source_path = args.source_path
+    roop.globals.target_path = args.target_path
+    roop.globals.output_path = util.normalize_output_path(roop.globals.source_path, roop.globals.target_path, args.output_path)
+    roop.globals.target_folder_path = args.target_folder_path
+    roop.globals.headless = args.source_path or args.target_path or args.output_path
+    # Always enable all processors when using GUI
+    if not roop.globals.headless:
+        roop.globals.frame_processors = ['face_swapper', 'face_enhancer']
+    else:
+        roop.globals.frame_processors = args.frame_processor
+
+    roop.globals.keep_fps = args.keep_fps
+    roop.globals.keep_frames = args.keep_frames
+    roop.globals.skip_audio = args.skip_audio
+    roop.globals.many_faces = args.many_faces
+    roop.globals.source_face_index = args.source_face_index
+    roop.globals.target_face_index = args.target_face_index
+    roop.globals.video_encoder = args.video_encoder
+    roop.globals.video_quality = args.video_quality
+    roop.globals.max_memory = args.max_memory
+    roop.globals.execution_providers = decode_execution_providers(args.execution_provider)
+    roop.globals.execution_threads = args.execution_threads
+
+
+def encode_execution_providers(execution_providers: List[str]) -> List[str]:
+    return [execution_provider.replace('ExecutionProvider', '').lower() for execution_provider in execution_providers]
+
+
+def decode_execution_providers(execution_providers: List[str]) -> List[str]:
+    return [provider for provider, encoded_execution_provider in zip(onnxruntime.get_available_providers(), encode_execution_providers(onnxruntime.get_available_providers()))
+            if any(execution_provider in encoded_execution_provider for execution_provider in execution_providers)]
+
+
+def suggest_max_memory() -> int:
+    if platform.system().lower() == 'darwin':
+        return 4
+    return 16
+
+
+def suggest_execution_providers() -> List[str]:
+    return encode_execution_providers(onnxruntime.get_available_providers())
+
+
+def suggest_execution_threads() -> int:
+    if 'DmlExecutionProvider' in roop.globals.execution_providers:
+        return 1
+    if 'ROCMExecutionProvider' in roop.globals.execution_providers:
+        return 1
+    return 8
+
+
+def limit_resources() -> None:
+    # prevent tensorflow memory leak
+    gpus = tensorflow.config.experimental.list_physical_devices('GPU')
+    for gpu in gpus:
+        tensorflow.config.experimental.set_virtual_device_configuration(gpu, [
+            tensorflow.config.experimental.VirtualDeviceConfiguration(memory_limit=1024)
+        ])
+    # limit memory usage
+    if roop.globals.max_memory:
+        memory = roop.globals.max_memory * 1024 ** 3
+        if platform.system().lower() == 'darwin':
+            memory = roop.globals.max_memory * 1024 ** 6
+        if platform.system().lower() == 'windows':
+            import ctypes
+            kernel32 = ctypes.windll.kernel32  # type: ignore[attr-defined]
+            kernel32.SetProcessWorkingSetSize(-1, ctypes.c_size_t(memory), ctypes.c_size_t(memory))
+        else:
+            import resource
+            resource.setrlimit(resource.RLIMIT_DATA, (memory, memory))
+
+
+
+def release_resources() -> None:
+    import gc
+    global process_mgr
+
+    if process_mgr is not None:
+        process_mgr.release_resources()
+        process_mgr = None
+
+    gc.collect()
+    if 'CUDAExecutionProvider' in roop.globals.execution_providers and torch.cuda.is_available():
+        with torch.cuda.device('cuda'):
+            torch.cuda.empty_cache()
+            torch.cuda.ipc_collect()
+
+
+def pre_check() -> bool:
+    if sys.version_info < (3, 9):
+        update_status('Python version is not supported - please upgrade to 3.9 or higher.')
+        return False
+    
+    download_directory_path = util.resolve_relative_path('../models')
+    util.conditional_download(download_directory_path, ['https://huggingface.co/countfloyd/deepfake/resolve/main/inswapper_128.onnx'])
+    util.conditional_download(download_directory_path, ['https://huggingface.co/countfloyd/deepfake/resolve/main/GFPGANv1.4.onnx'])
+    util.conditional_download(download_directory_path, ['https://github.com/csxmli2016/DMDNet/releases/download/v1/DMDNet.pth'])
+    download_directory_path = util.resolve_relative_path('../models/CLIP')
+    util.conditional_download(download_directory_path, ['https://huggingface.co/countfloyd/deepfake/resolve/main/rd64-uni-refined.pth'])
+    download_directory_path = util.resolve_relative_path('../models/CodeFormer')
+    util.conditional_download(download_directory_path, ['https://huggingface.co/countfloyd/deepfake/resolve/main/CodeFormerv0.1.onnx'])
+
+    if not shutil.which('ffmpeg'):
+       update_status('ffmpeg is not installed.')
+    return True
+
+def set_display_ui(function):
+    global call_display_ui
+
+    call_display_ui = function
+
+
+def update_status(message: str) -> None:
+    global call_display_ui
+
+    print(message)
+    if call_display_ui is not None:
+        call_display_ui(message)
+
+
+
+
+def start() -> None:
+    if roop.globals.headless:
+        faces = extract_face_images(roop.globals.source_path,  (False, 0))
+        roop.globals.INPUT_FACES.append(faces[roop.globals.source_face_index])
+        faces = extract_face_images(roop.globals.target_path,  (False, util.has_image_extension(roop.globals.target_path)))
+        roop.globals.TARGET_FACES.append(faces[roop.globals.target_face_index])
+        if 'face_enhancer' in roop.globals.frame_processors:
+            roop.globals.selected_enhancer = 'GFPGAN'
+       
+    batch_process(None, False, None)
+
+
+def get_processing_plugins(use_clip):
+    processors = "faceswap"
+    if use_clip:
+        processors += ",mask_clip2seg"
+    
+    if roop.globals.selected_enhancer == 'GFPGAN':
+        processors += ",gfpgan"
+    elif roop.globals.selected_enhancer == 'Codeformer':
+        processors += ",codeformer"
+    elif roop.globals.selected_enhancer == 'DMDNet':
+        processors += ",dmdnet"
+    return processors
+
+
+def live_swap(frame, swap_mode, use_clip, clip_text, selected_index = 0):
+    global process_mgr
+
+    if frame is None:
+        return frame
+
+    if process_mgr is None:
+        process_mgr = ProcessMgr(None)
+    
+    options = ProcessOptions(get_processing_plugins(use_clip), roop.globals.distance_threshold, roop.globals.blend_ratio, swap_mode, selected_index, clip_text)
+    process_mgr.initialize(roop.globals.INPUT_FACES, roop.globals.TARGET_FACES, options)
+    return process_mgr.process_frame(frame)
+
+
+def preview_mask(frame, clip_text):
+    import numpy as np
+    global process_mgr
+    
+    maskimage = np.zeros((frame.shape), np.uint8)
+    if process_mgr is None:
+        process_mgr = ProcessMgr()
+    options = ProcessOptions("mask_clip2seg", roop.globals.distance_threshold, roop.globals.blend_ratio, "None", 0, clip_text)
+    process_mgr.initialize(roop.globals.INPUT_FACES, roop.globals.TARGET_FACES, options)
+    return process_mgr.process_mask(frame, maskimage)
+    
+
+
+
+
+def batch_process(files:list[ProcessEntry], use_clip, new_clip_text, use_new_method, progress) -> None:
+    global clip_text, process_mgr
+
+    roop.globals.processing = True
+    release_resources()
+    limit_resources()
+
+    # limit threads for some providers
+    max_threads = suggest_execution_threads()
+    if max_threads == 1:
+        roop.globals.execution_threads = 1
+
+    imagefiles:list[ProcessEntry] = []
+    videofiles:list[ProcessEntry] = []
+           
+    update_status('Sorting videos/images')
+
+
+    for index, f in enumerate(files):
+        fullname = f.filename
+        if util.has_image_extension(fullname):
+            destination = util.get_destfilename_from_path(fullname, roop.globals.output_path, f'.{roop.globals.CFG.output_image_format}')
+            destination = util.replace_template(destination, index=index)
+            pathlib.Path(os.path.dirname(destination)).mkdir(parents=True, exist_ok=True)
+            f.finalname = destination
+            imagefiles.append(f)
+
+        elif util.is_video(fullname) or util.has_extension(fullname, ['gif']):
+            destination = util.get_destfilename_from_path(fullname, roop.globals.output_path, f'__temp.{roop.globals.CFG.output_video_format}')
+            f.finalname = destination
+            videofiles.append(f)
+
+
+    if process_mgr is None:
+        process_mgr = ProcessMgr(progress)
+    
+    options = ProcessOptions(get_processing_plugins(use_clip), roop.globals.distance_threshold, roop.globals.blend_ratio, roop.globals.face_swap_mode, 0, new_clip_text)
+    process_mgr.initialize(roop.globals.INPUT_FACES, roop.globals.TARGET_FACES, options)
+
+    if(len(imagefiles) > 0):
+        update_status('Processing image(s)')
+        origimages = []
+        fakeimages = []
+        for f in imagefiles:
+            origimages.append(f.filename)
+            fakeimages.append(f.finalname)
+
+        process_mgr.run_batch(origimages, fakeimages, roop.globals.execution_threads)
+        origimages.clear()
+        fakeimages.clear()
+
+    if(len(videofiles) > 0):
+        for index,v in enumerate(videofiles):
+            if not roop.globals.processing:
+                end_processing('Processing stopped!')
+                return
+            fps = v.fps if v.fps > 0 else util.detect_fps(v.filename)
+            update_status(f'Creating {os.path.basename(v.finalname)} with {fps} FPS...')
+            start_processing = time()
+            if roop.globals.keep_frames or not use_new_method:
+                util.create_temp(v.filename)
+                update_status('Extracting frames...')
+                util.extract_frames(v.filename,v.startframe,v.endframe, fps)
+                if not roop.globals.processing:
+                    end_processing('Processing stopped!')
+                    return
+
+                temp_frame_paths = util.get_temp_frame_paths(v.filename)
+                process_mgr.run_batch(temp_frame_paths, temp_frame_paths, roop.globals.execution_threads)
+                if not roop.globals.processing:
+                    end_processing('Processing stopped!')
+                    return
+                
+                util.create_video(v.filename, f.finalname, fps)
+                if not roop.globals.keep_frames:
+                    util.delete_temp_frames(temp_frame_paths[0])
+            else:
+                if util.has_extension(v.filename, ['gif']):
+                    skip_audio = True
+                else:
+                    skip_audio = roop.globals.skip_audio
+                process_mgr.run_batch_inmem(v.filename, v.finalname, v.startframe, v.endframe, fps,roop.globals.execution_threads, skip_audio)
+                
+            if not roop.globals.processing:
+                end_processing('Processing stopped!')
+                return
+            
+            video_file_name = v.finalname
+            if os.path.isfile(video_file_name):
+                destination = ''
+                if util.has_extension(v.filename, ['gif']):
+                    gifname = util.get_destfilename_from_path(v.filename, roop.globals.output_path, '.gif')
+                    destination = util.replace_template(gifname, index=index)
+                    pathlib.Path(os.path.dirname(destination)).mkdir(parents=True, exist_ok=True)
+
+                    update_status('Creating final GIF')
+                    util.create_gif_from_video(video_file_name, destination)
+                    if os.path.isfile(destination):
+                        os.remove(video_file_name)
+                else:
+                    skip_audio = roop.globals.skip_audio
+                    destination = util.replace_template(video_file_name, index=index)
+                    pathlib.Path(os.path.dirname(destination)).mkdir(parents=True, exist_ok=True)
+
+                    if not skip_audio:
+                        util.restore_audio(video_file_name, v.filename, v.startframe, v.endframe, destination)
+                        if os.path.isfile(destination):
+                            os.remove(video_file_name)
+                    else:
+                        shutil.move(video_file_name, destination)
+                update_status(f'\nProcessing {os.path.basename(destination)} took {time() - start_processing} secs')
+
+            else:
+                update_status(f'Failed processing {os.path.basename(v.finalname)}!')
+    end_processing('Finished')
+
+
+def end_processing(msg:str):
+    update_status(msg)
+    roop.globals.target_folder_path = None
+    release_resources()
+
+
+def destroy() -> None:
+    if roop.globals.target_path:
+        util.clean_temp(roop.globals.target_path)
+    release_resources()        
+    sys.exit()
+
+
+def run() -> None:
+    parse_args()
+    if not pre_check():
+        return
+    roop.globals.CFG = Settings('config.yaml')
+    if roop.globals.headless:
+        start()
+    else:
+        ui.run()

roop/face_util.py

@@ -0,0 +1,108 @@
+import threading
+from typing import Any
+import insightface
+
+import roop.globals
+from roop.typing import Frame, Face
+
+import cv2
+import numpy as np
+from roop.capturer import get_video_frame
+from roop.utilities import resolve_relative_path, conditional_download
+
+FACE_ANALYSER = None
+THREAD_LOCK_ANALYSER = threading.Lock()
+THREAD_LOCK_SWAPPER = threading.Lock()
+FACE_SWAPPER = None
+
+
+def get_face_analyser() -> Any:
+    global FACE_ANALYSER
+
+    with THREAD_LOCK_ANALYSER:
+        if FACE_ANALYSER is None:
+            if roop.globals.CFG.force_cpu:
+                print('Forcing CPU for Face Analysis')
+                FACE_ANALYSER = insightface.app.FaceAnalysis(name='buffalo_l', providers=['CPUExecutionProvider'])
+            else:
+                FACE_ANALYSER = insightface.app.FaceAnalysis(name='buffalo_l', providers=roop.globals.execution_providers)
+            FACE_ANALYSER.prepare(ctx_id=0, det_size=(640, 640) if roop.globals.default_det_size else (320,320))
+    return FACE_ANALYSER
+
+
+def get_first_face(frame: Frame) -> Any:
+    try:
+        faces = get_face_analyser().get(frame)
+        return min(faces, key=lambda x: x.bbox[0])
+    #   return sorted(faces, reverse=True, key=lambda x: (x.bbox[2] - x.bbox[0]) * (x.bbox[3] - x.bbox[1]))[0]
+    except:
+        return None
+
+
+def get_all_faces(frame: Frame) -> Any:
+    try:
+        faces = get_face_analyser().get(frame)
+        return sorted(faces, key = lambda x : x.bbox[0])
+    except:
+        return None
+
+
+def extract_face_images(source_filename, video_info):
+    face_data = []
+    source_image = None
+    
+    if video_info[0]:
+        frame = get_video_frame(source_filename, video_info[1])
+        if frame is not None:
+            source_image = frame
+        else:
+            return face_data
+    else:
+        source_image = cv2.imread(source_filename)
+
+        
+    faces = get_all_faces(source_image)
+    if faces is None:
+        return face_data
+
+    i = 0
+    for face in faces:
+        (startX, startY, endX, endY) = face['bbox'].astype("int")
+        face_temp = source_image[startY:endY, startX:endX]
+        if face_temp.size < 1:
+            continue
+        i += 1
+        face_data.append([face, face_temp])
+    return face_data
+
+
+
+
+def get_face_swapper() -> Any:
+    global FACE_SWAPPER
+
+    with THREAD_LOCK_SWAPPER:
+        if FACE_SWAPPER is None:
+            model_path = resolve_relative_path('../models/inswapper_128.onnx')
+            FACE_SWAPPER = insightface.model_zoo.get_model(model_path, providers=roop.globals.execution_providers)
+    return FACE_SWAPPER
+
+
+def pre_check() -> bool:
+    download_directory_path = resolve_relative_path('../models')
+    conditional_download(download_directory_path, ['https://huggingface.co/countfloyd/deepfake/resolve/main/inswapper_128.onnx'])
+    return True
+
+
+def swap_face(source_face: Face, target_face: Face, temp_frame: Frame) -> Frame:
+    return get_face_swapper().get(temp_frame, target_face, source_face, paste_back=True)
+
+def face_offset_top(face: Face, offset):
+    smallestmin = np.min(face.landmark_2d_106, 1)
+    smallest = smallestmin[1]
+    face['bbox'][1] += offset
+    face['bbox'][3] += offset
+    lm106 = face.landmark_2d_106
+    add = np.full_like(lm106, [0, offset])
+    face['landmark_2d_106'] = lm106 + add
+    return face

roop/ffmpeg_writer.py

@@ -0,0 +1,218 @@
+"""
+FFMPEG_Writer - write set of frames to video file
+
+original from
+https://github.com/Zulko/moviepy/blob/master/moviepy/video/io/ffmpeg_writer.py
+
+removed unnecessary dependencies
+
+The MIT License (MIT)
+
+Copyright (c) 2015 Zulko
+Copyright (c) 2023 Janvarev Vladislav
+"""
+
+import os
+import subprocess as sp
+
+PIPE = -1
+STDOUT = -2
+DEVNULL = -3
+
+FFMPEG_BINARY = "ffmpeg"
+
+class FFMPEG_VideoWriter:
+    """ A class for FFMPEG-based video writing.
+
+    A class to write videos using ffmpeg. ffmpeg will write in a large
+    choice of formats.
+
+    Parameters
+    -----------
+
+    filename
+      Any filename like 'video.mp4' etc. but if you want to avoid
+      complications it is recommended to use the generic extension
+      '.avi' for all your videos.
+
+    size
+      Size (width,height) of the output video in pixels.
+
+    fps
+      Frames per second in the output video file.
+
+    codec
+      FFMPEG codec. It seems that in terms of quality the hierarchy is
+      'rawvideo' = 'png' > 'mpeg4' > 'libx264'
+      'png' manages the same lossless quality as 'rawvideo' but yields
+      smaller files. Type ``ffmpeg -codecs`` in a terminal to get a list
+      of accepted codecs.
+
+      Note for default 'libx264': by default the pixel format yuv420p
+      is used. If the video dimensions are not both even (e.g. 720x405)
+      another pixel format is used, and this can cause problem in some
+      video readers.
+
+    audiofile
+      Optional: The name of an audio file that will be incorporated
+      to the video.
+
+    preset
+      Sets the time that FFMPEG will take to compress the video. The slower,
+      the better the compression rate. Possibilities are: ultrafast,superfast,
+      veryfast, faster, fast, medium (default), slow, slower, veryslow,
+      placebo.
+
+    bitrate
+      Only relevant for codecs which accept a bitrate. "5000k" offers
+      nice results in general.
+
+    """
+
+    def __init__(self, filename, size, fps, codec="libx265", crf=14, audiofile=None,
+                 preset="medium", bitrate=None,
+                 logfile=None, threads=None, ffmpeg_params=None):
+
+        if logfile is None:
+            logfile = sp.PIPE
+
+        self.filename = filename
+        self.codec = codec
+        self.ext = self.filename.split(".")[-1]
+        w = size[0] - 1 if size[0] % 2 != 0 else size[0]
+        h = size[1] - 1 if size[1] % 2 != 0 else size[1]
+
+
+        # order is important
+        cmd = [
+            FFMPEG_BINARY,
+            '-hide_banner',
+            '-hwaccel', 'auto',
+            '-y',
+            '-loglevel', 'error' if logfile == sp.PIPE else 'info',
+            '-f', 'rawvideo',
+            '-vcodec', 'rawvideo',
+            '-s', '%dx%d' % (size[0], size[1]),
+            #'-pix_fmt', 'rgba' if withmask else 'rgb24',
+            '-pix_fmt', 'bgr24',
+            '-r', str(fps),
+            '-an', '-i', '-' 
+        ]
+
+        if audiofile is not None:
+            cmd.extend([
+                '-i', audiofile,
+                '-acodec', 'copy'
+            ])
+
+        cmd.extend([
+            '-vcodec', codec,
+            '-crf', str(crf)
+            #'-preset', preset,
+        ])
+        if ffmpeg_params is not None:
+            cmd.extend(ffmpeg_params)
+        if bitrate is not None:
+            cmd.extend([
+                '-b', bitrate
+            ])
+
+        # scale to a resolution divisible by 2 if not even
+        cmd.extend(['-vf', f'scale={w}:{h}' if w != size[0] or h != size[1] else 'colorspace=bt709:iall=bt601-6-625:fast=1'])
+
+        if threads is not None:
+            cmd.extend(["-threads", str(threads)])
+
+        cmd.extend([
+            '-pix_fmt', 'yuv420p',
+
+        ])
+        cmd.extend([
+            filename
+        ])
+
+        test = str(cmd)
+        print(test)
+
+        popen_params = {"stdout": DEVNULL,
+                        "stderr": logfile,
+                        "stdin": sp.PIPE}
+
+        # This was added so that no extra unwanted window opens on windows
+        # when the child process is created
+        if os.name == "nt":
+            popen_params["creationflags"] = 0x08000000  # CREATE_NO_WINDOW
+        
+        self.proc = sp.Popen(cmd, **popen_params)
+
+
+    def write_frame(self, img_array):
+        """ Writes one frame in the file."""
+        try:
+            #if PY3:
+            self.proc.stdin.write(img_array.tobytes())
+            # else:
+            #    self.proc.stdin.write(img_array.tostring())
+        except IOError as err:
+            _, ffmpeg_error = self.proc.communicate()
+            error = (str(err) + ("\n\nroop unleashed error: FFMPEG encountered "
+                                 "the following error while writing file %s:"
+                                 "\n\n %s" % (self.filename, str(ffmpeg_error))))
+
+            if b"Unknown encoder" in ffmpeg_error:
+
+                error = error+("\n\nThe video export "
+                  "failed because FFMPEG didn't find the specified "
+                  "codec for video encoding (%s). Please install "
+                  "this codec or change the codec when calling "
+                  "write_videofile. For instance:\n"
+                  "  >>> clip.write_videofile('myvid.webm', codec='libvpx')")%(self.codec)
+
+            elif b"incorrect codec parameters ?" in ffmpeg_error:
+
+                 error = error+("\n\nThe video export "
+                  "failed, possibly because the codec specified for "
+                  "the video (%s) is not compatible with the given "
+                  "extension (%s). Please specify a valid 'codec' "
+                  "argument in write_videofile. This would be 'libx264' "
+                  "or 'mpeg4' for mp4, 'libtheora' for ogv, 'libvpx for webm. "
+                  "Another possible reason is that the audio codec was not "
+                  "compatible with the video codec. For instance the video "
+                  "extensions 'ogv' and 'webm' only allow 'libvorbis' (default) as a"
+                  "video codec."
+                  )%(self.codec, self.ext)
+
+            elif  b"encoder setup failed" in ffmpeg_error:
+
+                error = error+("\n\nThe video export "
+                  "failed, possibly because the bitrate you specified "
+                  "was too high or too low for the video codec.")
+
+            elif b"Invalid encoder type" in ffmpeg_error:
+
+                error = error + ("\n\nThe video export failed because the codec "
+                  "or file extension you provided is not a video")
+
+
+            raise IOError(error)
+
+    def close(self):
+        if self.proc:
+            self.proc.stdin.close()
+            if self.proc.stderr is not None:
+                self.proc.stderr.close()
+            self.proc.wait()
+
+        self.proc = None
+
+    # Support the Context Manager protocol, to ensure that resources are cleaned up.
+
+    def __enter__(self):
+        return self
+
+    def __exit__(self, exc_type, exc_value, traceback):
+        self.close()
+
+
+
+    

roop/globals.py

@@ -0,0 +1,44 @@
+from settings import Settings
+from typing import List
+
+source_path = None
+target_path = None
+output_path = None
+target_folder_path = None
+
+frame_processors: List[str] = []
+keep_fps = None
+keep_frames = None
+skip_audio = None
+many_faces = None
+use_batch = None
+source_face_index = 0
+target_face_index = 0
+face_position = None
+video_encoder = None
+video_quality = None
+max_memory = None
+execution_providers: List[str] = []
+execution_threads = None
+headless = None
+log_level = 'error'
+selected_enhancer = None
+face_swap_mode = None
+blend_ratio = 0.5
+distance_threshold = 0.65
+default_det_size = True
+
+processing = False 
+
+FACE_ENHANCER = None
+
+INPUT_FACES = []
+TARGET_FACES = []
+
+IMAGE_CHAIN_PROCESSOR = None
+VIDEO_CHAIN_PROCESSOR = None
+BATCH_IMAGE_CHAIN_PROCESSOR = None
+
+CFG: Settings = None
+
+

roop/metadata.py

@@ -0,0 +1,2 @@
+name = 'roop unleashed'
+version = '3.0.2'

roop/processors/Enhance_Codeformer.py

@@ -0,0 +1,70 @@
+from typing import Any, List, Callable
+import cv2 
+import threading
+import numpy as np
+import onnxruntime
+import onnx
+import roop.globals
+
+from roop.typing import Face, Frame
+from roop.utilities import resolve_relative_path
+
+
+# THREAD_LOCK = threading.Lock()
+
+
+class Enhance_CodeFormer():
+    model_codeformer = None
+    devicename = None
+
+    processorname = 'codeformer'
+    type = 'enhance'
+    
+
+    def Initialize(self, devicename):
+        if self.model_codeformer is None:
+            self.devicename = devicename
+            model_path = resolve_relative_path('../models/CodeFormer/CodeFormerv0.1.onnx')
+            self.model_codeformer = onnxruntime.InferenceSession(model_path, None, providers=roop.globals.execution_providers)
+            self.model_inputs = self.model_codeformer.get_inputs()
+            model_outputs = self.model_codeformer.get_outputs()
+            self.io_binding = self.model_codeformer.io_binding()           
+            self.io_binding.bind_cpu_input(self.model_inputs[1].name, np.array([0.5]))
+            self.io_binding.bind_output(model_outputs[0].name, self.devicename)
+
+
+    def Run(self, source_face: Face, target_face: Face, temp_frame: Frame) -> Frame:
+        input_size = temp_frame.shape[1]
+        # preprocess
+        temp_frame = cv2.resize(temp_frame, (512, 512), cv2.INTER_CUBIC)
+        temp_frame = cv2.cvtColor(temp_frame, cv2.COLOR_BGR2RGB)
+        temp_frame = temp_frame.astype('float32') / 255.0
+        temp_frame = (temp_frame - 0.5) / 0.5
+        temp_frame = np.expand_dims(temp_frame, axis=0).transpose(0, 3, 1, 2)
+        
+        self.io_binding.bind_cpu_input(self.model_inputs[0].name, temp_frame.astype(np.float32))
+        self.model_codeformer.run_with_iobinding(self.io_binding)
+        ort_outs = self.io_binding.copy_outputs_to_cpu()
+        result = ort_outs[0][0]
+        del ort_outs
+        
+        # post-process
+        result = result.transpose((1, 2, 0))
+
+        un_min = -1.0
+        un_max = 1.0
+        result = np.clip(result, un_min, un_max)
+        result = (result - un_min) / (un_max - un_min)
+
+        result = cv2.cvtColor(result, cv2.COLOR_RGB2BGR)
+        result = (result * 255.0).round()
+        scale_factor = int(result.shape[1] / input_size)       
+        return result.astype(np.uint8), scale_factor
+
+
+    def Release(self):
+        del self.model_codeformer
+        self.model_codeformer = None
+        del self.io_binding
+        self.io_binding = None
+

roop/processors/Enhance_DMDNet.py

@@ -0,0 +1,861 @@
+from typing import Any, List, Callable
+import cv2 
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.nn.utils.spectral_norm as SpectralNorm
+import threading
+from torchvision.ops import roi_align
+
+from math import sqrt
+
+from torchvision.transforms.functional import normalize
+import roop.globals
+
+from roop.typing import Face, Frame
+from roop.utilities import resolve_relative_path
+
+
+THREAD_LOCK_DMDNET = threading.Lock()
+
+
+class Enhance_DMDNet():
+
+    model_dmdnet = None
+    torchdevice = None
+
+    processorname = 'dmdnet'
+    type = 'enhance'
+
+
+    def Initialize(self, devicename):
+        if self.model_dmdnet is None:
+            self.model_dmdnet = self.create(devicename)
+            
+
+    # temp_frame already cropped+aligned, bbox not
+    def Run(self, source_face: Face, target_face: Face, temp_frame: Frame) -> Frame:
+        input_size = temp_frame.shape[1]
+
+        result = self.enhance_face(source_face, temp_frame, target_face)
+        scale_factor = int(result.shape[1] / input_size)       
+        return result.astype(np.uint8), scale_factor
+
+
+    def Release(self):
+        self.model_gfpgan = None
+
+
+    # https://stackoverflow.com/a/67174339
+    def landmarks106_to_68(self, pt106):
+        map106to68=[1,10,12,14,16,3,5,7,0,23,21,19,32,30,28,26,17,
+                        43,48,49,51,50,
+                        102,103,104,105,101,
+                        72,73,74,86,78,79,80,85,84,
+                        35,41,42,39,37,36,
+                        89,95,96,93,91,90,
+                        52,64,63,71,67,68,61,58,59,53,56,55,65,66,62,70,69,57,60,54
+                        ]
+      
+        pt68 = []
+        for i in range(68):
+            index = map106to68[i]
+            pt68.append(pt106[index])
+        return pt68
+
+        
+
+
+    def check_bbox(self, imgs, boxes):
+        boxes = boxes.view(-1, 4, 4)
+        colors = [(0, 255, 0), (0, 255, 0), (255, 255, 0), (255, 0, 0)]
+        i = 0
+        for img, box in zip(imgs, boxes):
+            img = (img + 1)/2 * 255
+            img2 = img.permute(1, 2, 0).float().cpu().flip(2).numpy().copy()
+            for idx, point in enumerate(box):
+                cv2.rectangle(img2, (int(point[0]), int(point[1])), (int(point[2]), int(point[3])), color=colors[idx], thickness=2)
+            cv2.imwrite('dmdnet_{:02d}.png'.format(i), img2)
+            i += 1
+
+
+    def trans_points2d(self, pts, M):
+        new_pts = np.zeros(shape=pts.shape, dtype=np.float32)
+        for i in range(pts.shape[0]):
+            pt = pts[i]
+            new_pt = np.array([pt[0], pt[1], 1.0], dtype=np.float32)
+            new_pt = np.dot(M, new_pt)
+            new_pts[i] = new_pt[0:2]
+
+        return new_pts
+
+
+    def enhance_face(self, ref_face, temp_frame, face):
+        # preprocess
+        start_x, start_y, end_x, end_y = map(int, face['bbox'])
+        width = end_x - start_x
+        height = end_y - start_y
+        
+        
+        lm106 = face.landmark_2d_106
+        lq_landmarks = np.asarray(self.landmarks106_to_68(lm106))
+
+        if temp_frame.shape[0] != 512 or temp_frame.shape[1] != 512:
+            # scale to 512x512
+            scale_factor = 512 / temp_frame.shape[1]
+
+            M = face.matrix * scale_factor
+
+            lq_landmarks = self.trans_points2d(lq_landmarks, M)
+            temp_frame = cv2.resize(temp_frame, (512,512), interpolation = cv2.INTER_AREA)
+
+        if temp_frame.ndim == 2:
+            temp_frame = cv2.cvtColor(temp_frame, cv2.COLOR_GRAY2RGB)  # GGG
+        # else:
+        #     temp_frame = cv2.cvtColor(temp_frame, cv2.COLOR_BGR2RGB)  # RGB
+
+        lq = read_img_tensor(temp_frame)
+
+        LQLocs = get_component_location(lq_landmarks)
+        # self.check_bbox(lq, LQLocs.unsqueeze(0))
+
+        # specific
+        # start_x, start_y, end_x, end_y = map(int, ref_face['bbox'])
+        # temp_face = temp_frame[start_y:end_y, start_x:end_x]
+        # if temp_face.size:
+        #     SpecificImgs = []
+        #     SpecificLocs = []
+        #     lm106 = ref_face.landmark_2d_106
+        #     ref_landmarks = asarray(self.landmarks106_to_68(lm106))
+        #     ref_tensor, ref_landmarks, _ = read_img_tensor(clip, ref_landmarks)
+        #     SpecificImgs.append(ref_tensor)
+        #     ref_locs = get_component_location(ref_landmarks)
+        #     SpecificLocs.append(ref_locs.unsqueeze(0))
+
+        #     SpecificImgs = torch.cat(SpecificImgs, dim=0)
+        #     SpecificLocs = torch.cat(SpecificLocs, dim=0)
+        #     # check_bbox(SpecificImgs, SpecificLocs)
+        #     SpMem256, SpMem128, SpMem64 = DMDNet.generate_specific_dictionary(sp_imgs = SpecificImgs.to(device), sp_locs = SpecificLocs)
+        #     SpMem256Para = {}
+        #     SpMem128Para = {}
+        #     SpMem64Para = {}
+        #     for k, v in SpMem256.items():
+        #         SpMem256Para[k] = v
+        #     for k, v in SpMem128.items():
+        #         SpMem128Para[k] = v
+        #     for k, v in SpMem64.items():
+        #         SpMem64Para[k] = v
+
+        # generic
+        SpMem256Para, SpMem128Para, SpMem64Para = None, None, None
+
+        with torch.no_grad():
+            with THREAD_LOCK_DMDNET:
+                try:
+                    GenericResult, SpecificResult = self.model_dmdnet(lq = lq.to(self.torchdevice), loc = LQLocs.unsqueeze(0), sp_256 = SpMem256Para, sp_128 = SpMem128Para, sp_64 = SpMem64Para)
+                except Exception as e:
+                    print(f'Error {e} there may be something wrong with the detected component locations.')
+                    return temp_frame
+        save_generic = GenericResult * 0.5 + 0.5
+        save_generic = save_generic.squeeze(0).permute(1, 2, 0).flip(2) # RGB->BGR
+        save_generic = np.clip(save_generic.float().cpu().numpy(), 0, 1) * 255.0
+
+        check_lq = lq * 0.5 + 0.5
+        check_lq = check_lq.squeeze(0).permute(1, 2, 0).flip(2) # RGB->BGR
+        check_lq = np.clip(check_lq.float().cpu().numpy(), 0, 1) * 255.0
+        enhanced_img =  np.hstack((check_lq, save_generic))
+        temp_frame =  save_generic.astype("uint8")
+        temp_frame = cv2.cvtColor(temp_frame, cv2.COLOR_RGB2BGR)  # RGB
+        return temp_frame
+    
+
+    def create(self, devicename):
+        self.torchdevice = torch.device(devicename)
+        model_dmdnet = DMDNet().to(self.torchdevice)
+        weights = torch.load('./models/DMDNet.pth') 
+        model_dmdnet.load_state_dict(weights, strict=True)
+
+        model_dmdnet.eval()
+        num_params = 0
+        for param in model_dmdnet.parameters():
+            num_params += param.numel()
+        return model_dmdnet
+
+    # print('{:>8s} : {}'.format('Using device', device))
+    # print('{:>8s} : {:.2f}M'.format('Model params', num_params/1e6))
+
+
+
+def read_img_tensor(Img=None): #rgb -1~1 
+    Img = Img.transpose((2, 0, 1))/255.0
+    Img = torch.from_numpy(Img).float()
+    normalize(Img, [0.5,0.5,0.5], [0.5,0.5,0.5], inplace=True)
+    ImgTensor = Img.unsqueeze(0)
+    return ImgTensor
+
+
+def get_component_location(Landmarks, re_read=False):
+    if re_read:
+        ReadLandmark = []
+        with open(Landmarks,'r') as f:
+            for line in f:
+                tmp = [float(i) for i in line.split(' ') if i != '\n']
+                ReadLandmark.append(tmp)
+        ReadLandmark = np.array(ReadLandmark) #
+        Landmarks = np.reshape(ReadLandmark, [-1, 2]) # 68*2
+    Map_LE_B = list(np.hstack((range(17,22), range(36,42))))
+    Map_RE_B = list(np.hstack((range(22,27), range(42,48))))
+    Map_LE = list(range(36,42))
+    Map_RE = list(range(42,48))
+    Map_NO = list(range(29,36))
+    Map_MO = list(range(48,68))
+
+    Landmarks[Landmarks>504]=504
+    Landmarks[Landmarks<8]=8
+    
+    #left eye
+    Mean_LE = np.mean(Landmarks[Map_LE],0)
+    L_LE1 = Mean_LE[1] - np.min(Landmarks[Map_LE_B,1])
+    L_LE1 = L_LE1 * 1.3
+    L_LE2 = L_LE1 / 1.9
+    L_LE_xy = L_LE1 + L_LE2
+    L_LE_lt = [L_LE_xy/2, L_LE1]
+    L_LE_rb = [L_LE_xy/2, L_LE2]
+    Location_LE = np.hstack((Mean_LE - L_LE_lt + 1, Mean_LE + L_LE_rb)).astype(int)
+
+    #right eye
+    Mean_RE = np.mean(Landmarks[Map_RE],0)
+    L_RE1 = Mean_RE[1] - np.min(Landmarks[Map_RE_B,1])
+    L_RE1 = L_RE1 * 1.3
+    L_RE2 = L_RE1 / 1.9
+    L_RE_xy = L_RE1 + L_RE2
+    L_RE_lt = [L_RE_xy/2, L_RE1]
+    L_RE_rb = [L_RE_xy/2, L_RE2]
+    Location_RE = np.hstack((Mean_RE - L_RE_lt + 1, Mean_RE + L_RE_rb)).astype(int)
+
+    #nose
+    Mean_NO = np.mean(Landmarks[Map_NO],0)
+    L_NO1 =( np.max([Mean_NO[0] - Landmarks[31][0], Landmarks[35][0] - Mean_NO[0]])) * 1.25
+    L_NO2 = (Landmarks[33][1] - Mean_NO[1]) * 1.1
+    L_NO_xy = L_NO1 * 2
+    L_NO_lt = [L_NO_xy/2, L_NO_xy - L_NO2]
+    L_NO_rb = [L_NO_xy/2, L_NO2]
+    Location_NO = np.hstack((Mean_NO - L_NO_lt + 1, Mean_NO + L_NO_rb)).astype(int)
+    
+    #mouth
+    Mean_MO = np.mean(Landmarks[Map_MO],0)
+    L_MO = np.max((np.max(np.max(Landmarks[Map_MO],0) - np.min(Landmarks[Map_MO],0))/2,16)) * 1.1
+    MO_O = Mean_MO - L_MO + 1
+    MO_T = Mean_MO + L_MO
+    MO_T[MO_T>510]=510
+    Location_MO = np.hstack((MO_O, MO_T)).astype(int)
+    return torch.cat([torch.FloatTensor(Location_LE).unsqueeze(0), torch.FloatTensor(Location_RE).unsqueeze(0), torch.FloatTensor(Location_NO).unsqueeze(0), torch.FloatTensor(Location_MO).unsqueeze(0)], dim=0)
+
+
+
+
+def calc_mean_std_4D(feat, eps=1e-5):
+    # eps is a small value added to the variance to avoid divide-by-zero.
+    size = feat.size()
+    assert (len(size) == 4)
+    N, C = size[:2]
+    feat_var = feat.view(N, C, -1).var(dim=2) + eps
+    feat_std = feat_var.sqrt().view(N, C, 1, 1)
+    feat_mean = feat.view(N, C, -1).mean(dim=2).view(N, C, 1, 1)
+    return feat_mean, feat_std
+
+def adaptive_instance_normalization_4D(content_feat, style_feat): # content_feat is ref feature, style is degradate feature
+    size = content_feat.size()
+    style_mean, style_std = calc_mean_std_4D(style_feat)
+    content_mean, content_std = calc_mean_std_4D(content_feat)
+    normalized_feat = (content_feat - content_mean.expand(size)) / content_std.expand(size)
+    return normalized_feat * style_std.expand(size) + style_mean.expand(size)
+
+
+def convU(in_channels, out_channels,conv_layer, norm_layer, kernel_size=3, stride=1,dilation=1, bias=True):
+    return nn.Sequential(
+        SpectralNorm(conv_layer(in_channels, out_channels, kernel_size=kernel_size, stride=stride, dilation=dilation, padding=((kernel_size-1)//2)*dilation, bias=bias)),
+        nn.LeakyReLU(0.2),
+        SpectralNorm(conv_layer(out_channels, out_channels, kernel_size=kernel_size, stride=stride, dilation=dilation, padding=((kernel_size-1)//2)*dilation, bias=bias)),
+    )
+    
+
+class MSDilateBlock(nn.Module):
+    def __init__(self, in_channels,conv_layer=nn.Conv2d, norm_layer=nn.BatchNorm2d, kernel_size=3, dilation=[1,1,1,1], bias=True):
+        super(MSDilateBlock, self).__init__()
+        self.conv1 =  convU(in_channels, in_channels,conv_layer, norm_layer, kernel_size,dilation=dilation[0], bias=bias)
+        self.conv2 =  convU(in_channels, in_channels,conv_layer, norm_layer, kernel_size,dilation=dilation[1], bias=bias)
+        self.conv3 =  convU(in_channels, in_channels,conv_layer, norm_layer, kernel_size,dilation=dilation[2], bias=bias)
+        self.conv4 =  convU(in_channels, in_channels,conv_layer, norm_layer, kernel_size,dilation=dilation[3], bias=bias)
+        self.convi =  SpectralNorm(conv_layer(in_channels*4, in_channels, kernel_size=kernel_size, stride=1, padding=(kernel_size-1)//2, bias=bias))
+    def forward(self, x):
+        conv1 = self.conv1(x)
+        conv2 = self.conv2(x)
+        conv3 = self.conv3(x)
+        conv4 = self.conv4(x)
+        cat  = torch.cat([conv1, conv2, conv3, conv4], 1)
+        out = self.convi(cat) + x
+        return out
+
+
+class AdaptiveInstanceNorm(nn.Module):
+    def __init__(self, in_channel):
+        super().__init__()
+        self.norm = nn.InstanceNorm2d(in_channel)
+
+    def forward(self, input, style):
+        style_mean, style_std = calc_mean_std_4D(style)
+        out = self.norm(input)
+        size = input.size()
+        out = style_std.expand(size) * out + style_mean.expand(size)
+        return out
+
+class NoiseInjection(nn.Module):
+    def __init__(self, channel):
+        super().__init__()
+        self.weight = nn.Parameter(torch.zeros(1, channel, 1, 1))
+    def forward(self, image, noise):
+        if noise is None:
+            b, c, h, w = image.shape
+            noise = image.new_empty(b, 1, h, w).normal_()
+        return image + self.weight * noise
+
+class StyledUpBlock(nn.Module):
+    def __init__(self, in_channel, out_channel, kernel_size=3, padding=1,upsample=False, noise_inject=False):
+        super().__init__()
+
+        self.noise_inject = noise_inject
+        if upsample:
+            self.conv1 = nn.Sequential(
+                nn.Upsample(scale_factor=2, mode='bilinear', align_corners=False),
+                SpectralNorm(nn.Conv2d(in_channel, out_channel, kernel_size, padding=padding)),
+                nn.LeakyReLU(0.2),
+            )
+        else:
+            self.conv1 = nn.Sequential(
+                SpectralNorm(nn.Conv2d(in_channel, out_channel, kernel_size, padding=padding)),
+                nn.LeakyReLU(0.2),
+                SpectralNorm(nn.Conv2d(out_channel, out_channel, kernel_size, padding=padding)),
+            )
+        self.convup = nn.Sequential(
+                nn.Upsample(scale_factor=2, mode='bilinear', align_corners=False),
+                SpectralNorm(nn.Conv2d(out_channel, out_channel, kernel_size, padding=padding)),
+                nn.LeakyReLU(0.2),
+                SpectralNorm(nn.Conv2d(out_channel, out_channel, kernel_size, padding=padding)),
+            )
+        if self.noise_inject:
+            self.noise1 = NoiseInjection(out_channel)
+
+        self.lrelu1 = nn.LeakyReLU(0.2)
+
+        self.ScaleModel1 = nn.Sequential(
+            SpectralNorm(nn.Conv2d(in_channel,out_channel,3, 1, 1)),
+            nn.LeakyReLU(0.2),
+            SpectralNorm(nn.Conv2d(out_channel, out_channel, 3, 1, 1))
+        )
+        self.ShiftModel1 = nn.Sequential(
+            SpectralNorm(nn.Conv2d(in_channel,out_channel,3, 1, 1)),
+            nn.LeakyReLU(0.2),
+            SpectralNorm(nn.Conv2d(out_channel, out_channel, 3, 1, 1)),
+        )
+       
+    def forward(self, input, style):
+        out = self.conv1(input)
+        out = self.lrelu1(out)
+        Shift1 = self.ShiftModel1(style)
+        Scale1 = self.ScaleModel1(style)
+        out = out * Scale1 + Shift1
+        if self.noise_inject:
+            out = self.noise1(out, noise=None)
+        outup = self.convup(out)
+        return outup
+
+
+####################################################################
+###############Face Dictionary Generator
+####################################################################
+def AttentionBlock(in_channel):
+    return nn.Sequential(
+        SpectralNorm(nn.Conv2d(in_channel, in_channel, 3, 1, 1)),
+        nn.LeakyReLU(0.2),
+        SpectralNorm(nn.Conv2d(in_channel, in_channel, 3, 1, 1)),
+    )
+
+class DilateResBlock(nn.Module):
+    def __init__(self, dim, dilation=[5,3] ):
+        super(DilateResBlock, self).__init__()
+        self.Res = nn.Sequential(
+            SpectralNorm(nn.Conv2d(dim, dim, 3, 1, ((3-1)//2)*dilation[0], dilation[0])),
+            nn.LeakyReLU(0.2),
+            SpectralNorm(nn.Conv2d(dim, dim, 3, 1, ((3-1)//2)*dilation[1], dilation[1])),
+        )
+    def forward(self, x):
+        out = x + self.Res(x)
+        return out
+
+
+class KeyValue(nn.Module):
+    def __init__(self, indim, keydim, valdim):
+        super(KeyValue, self).__init__()
+        self.Key = nn.Sequential(
+            SpectralNorm(nn.Conv2d(indim, keydim, kernel_size=(3,3), padding=(1,1), stride=1)),
+            nn.LeakyReLU(0.2),
+            SpectralNorm(nn.Conv2d(keydim, keydim, kernel_size=(3,3), padding=(1,1), stride=1)),
+        )
+        self.Value = nn.Sequential(
+            SpectralNorm(nn.Conv2d(indim, valdim, kernel_size=(3,3), padding=(1,1), stride=1)),
+            nn.LeakyReLU(0.2),
+            SpectralNorm(nn.Conv2d(valdim, valdim, kernel_size=(3,3), padding=(1,1), stride=1)),
+        )
+    def forward(self, x):  
+        return self.Key(x), self.Value(x)
+
+class MaskAttention(nn.Module):
+    def __init__(self, indim):
+        super(MaskAttention, self).__init__()
+        self.conv1 = nn.Sequential(
+            SpectralNorm(nn.Conv2d(indim, indim//3, kernel_size=(3,3), padding=(1,1), stride=1)),
+            nn.LeakyReLU(0.2),
+            SpectralNorm(nn.Conv2d(indim//3, indim//3, kernel_size=(3,3), padding=(1,1), stride=1)),
+        )
+        self.conv2 = nn.Sequential(
+            SpectralNorm(nn.Conv2d(indim, indim//3, kernel_size=(3,3), padding=(1,1), stride=1)),
+            nn.LeakyReLU(0.2),
+            SpectralNorm(nn.Conv2d(indim//3, indim//3, kernel_size=(3,3), padding=(1,1), stride=1)),
+        )
+        self.conv3 = nn.Sequential(
+            SpectralNorm(nn.Conv2d(indim, indim//3, kernel_size=(3,3), padding=(1,1), stride=1)),
+            nn.LeakyReLU(0.2),
+            SpectralNorm(nn.Conv2d(indim//3, indim//3, kernel_size=(3,3), padding=(1,1), stride=1)),
+        )
+        self.convCat = nn.Sequential(
+            SpectralNorm(nn.Conv2d(indim//3 * 3, indim, kernel_size=(3,3), padding=(1,1), stride=1)),
+            nn.LeakyReLU(0.2),
+            SpectralNorm(nn.Conv2d(indim, indim, kernel_size=(3,3), padding=(1,1), stride=1)),
+        )
+    def forward(self, x, y, z):
+        c1 = self.conv1(x)
+        c2 = self.conv2(y)
+        c3 = self.conv3(z)
+        return self.convCat(torch.cat([c1,c2,c3], dim=1))
+
+class Query(nn.Module):
+    def __init__(self, indim, quedim):
+        super(Query, self).__init__()
+        self.Query = nn.Sequential(
+            SpectralNorm(nn.Conv2d(indim, quedim, kernel_size=(3,3), padding=(1,1), stride=1)),
+            nn.LeakyReLU(0.2),
+            SpectralNorm(nn.Conv2d(quedim, quedim, kernel_size=(3,3), padding=(1,1), stride=1)),
+        )
+    def forward(self, x):
+        return self.Query(x)
+
+def roi_align_self(input, location, target_size):
+    return torch.cat([F.interpolate(input[i:i+1,:,location[i,1]:location[i,3],location[i,0]:location[i,2]],(target_size,target_size),mode='bilinear',align_corners=False) for i in range(input.size(0))],0)
+
+class FeatureExtractor(nn.Module):
+    def __init__(self, ngf = 64, key_scale = 4):#
+        super().__init__()
+
+        self.key_scale = 4
+        self.part_sizes = np.array([80,80,50,110]) #
+        self.feature_sizes = np.array([256,128,64]) # 
+
+        self.conv1 = nn.Sequential(
+                SpectralNorm(nn.Conv2d(3, ngf, 3, 2, 1)),
+                nn.LeakyReLU(0.2),
+                SpectralNorm(nn.Conv2d(ngf, ngf, 3, 1, 1)),
+            )
+        self.conv2 = nn.Sequential(
+            SpectralNorm(nn.Conv2d(ngf, ngf, 3, 1, 1)),
+            nn.LeakyReLU(0.2),
+            SpectralNorm(nn.Conv2d(ngf, ngf, 3, 1, 1))
+        )
+        self.res1 = DilateResBlock(ngf, [5,3])
+        self.res2 = DilateResBlock(ngf, [5,3])
+
+        
+        self.conv3 = nn.Sequential(
+            SpectralNorm(nn.Conv2d(ngf, ngf*2, 3, 2, 1)),
+            nn.LeakyReLU(0.2),
+            SpectralNorm(nn.Conv2d(ngf*2, ngf*2, 3, 1, 1)),
+            )
+        self.conv4 = nn.Sequential(
+            SpectralNorm(nn.Conv2d(ngf*2, ngf*2, 3, 1, 1)),
+            nn.LeakyReLU(0.2),
+            SpectralNorm(nn.Conv2d(ngf*2, ngf*2, 3, 1, 1))
+        )
+        self.res3 = DilateResBlock(ngf*2, [3,1])
+        self.res4 = DilateResBlock(ngf*2, [3,1])
+
+        self.conv5 = nn.Sequential(
+            SpectralNorm(nn.Conv2d(ngf*2, ngf*4, 3, 2, 1)),
+            nn.LeakyReLU(0.2),
+            SpectralNorm(nn.Conv2d(ngf*4, ngf*4, 3, 1, 1)),
+        )
+        self.conv6 = nn.Sequential(
+            SpectralNorm(nn.Conv2d(ngf*4, ngf*4, 3, 1, 1)),
+            nn.LeakyReLU(0.2),
+            SpectralNorm(nn.Conv2d(ngf*4, ngf*4, 3, 1, 1))
+        )
+        self.res5 = DilateResBlock(ngf*4, [1,1])
+        self.res6 = DilateResBlock(ngf*4, [1,1])
+
+        self.LE_256_Q = Query(ngf, ngf // self.key_scale)
+        self.RE_256_Q = Query(ngf, ngf // self.key_scale)
+        self.MO_256_Q = Query(ngf, ngf // self.key_scale)
+        self.LE_128_Q = Query(ngf * 2, ngf * 2 // self.key_scale)
+        self.RE_128_Q = Query(ngf * 2, ngf * 2 // self.key_scale)
+        self.MO_128_Q = Query(ngf * 2, ngf * 2 // self.key_scale)
+        self.LE_64_Q = Query(ngf * 4, ngf * 4 // self.key_scale)
+        self.RE_64_Q = Query(ngf * 4, ngf * 4 // self.key_scale)
+        self.MO_64_Q = Query(ngf * 4, ngf * 4 // self.key_scale)
+
+
+    def forward(self, img, locs):
+        le_location = locs[:,0,:].int().cpu().numpy()
+        re_location = locs[:,1,:].int().cpu().numpy()
+        no_location = locs[:,2,:].int().cpu().numpy()
+        mo_location = locs[:,3,:].int().cpu().numpy()
+        
+
+        f1_0 = self.conv1(img) 
+        f1_1 = self.res1(f1_0)
+        f2_0 = self.conv2(f1_1)
+        f2_1 = self.res2(f2_0)
+
+        f3_0 = self.conv3(f2_1) 
+        f3_1 = self.res3(f3_0)
+        f4_0 = self.conv4(f3_1)
+        f4_1 = self.res4(f4_0)
+
+        f5_0 = self.conv5(f4_1) 
+        f5_1 = self.res5(f5_0)
+        f6_0 = self.conv6(f5_1)
+        f6_1 = self.res6(f6_0)
+
+
+        ####ROI Align
+        le_part_256 = roi_align_self(f2_1.clone(), le_location//2, self.part_sizes[0]//2)
+        re_part_256 = roi_align_self(f2_1.clone(), re_location//2, self.part_sizes[1]//2)
+        mo_part_256 = roi_align_self(f2_1.clone(), mo_location//2, self.part_sizes[3]//2)
+
+        le_part_128 = roi_align_self(f4_1.clone(), le_location//4, self.part_sizes[0]//4)
+        re_part_128 = roi_align_self(f4_1.clone(), re_location//4, self.part_sizes[1]//4)
+        mo_part_128 = roi_align_self(f4_1.clone(), mo_location//4, self.part_sizes[3]//4)
+
+        le_part_64 = roi_align_self(f6_1.clone(), le_location//8, self.part_sizes[0]//8)
+        re_part_64 = roi_align_self(f6_1.clone(), re_location//8, self.part_sizes[1]//8)
+        mo_part_64 = roi_align_self(f6_1.clone(), mo_location//8, self.part_sizes[3]//8)
+
+
+        le_256_q = self.LE_256_Q(le_part_256)
+        re_256_q = self.RE_256_Q(re_part_256)
+        mo_256_q = self.MO_256_Q(mo_part_256)
+
+        le_128_q = self.LE_128_Q(le_part_128)
+        re_128_q = self.RE_128_Q(re_part_128)
+        mo_128_q = self.MO_128_Q(mo_part_128)
+
+        le_64_q = self.LE_64_Q(le_part_64)
+        re_64_q = self.RE_64_Q(re_part_64)
+        mo_64_q = self.MO_64_Q(mo_part_64)
+
+        return {'f256': f2_1, 'f128': f4_1, 'f64': f6_1,\
+            'le256': le_part_256, 're256': re_part_256, 'mo256': mo_part_256, \
+            'le128': le_part_128, 're128': re_part_128, 'mo128': mo_part_128, \
+            'le64': le_part_64, 're64': re_part_64, 'mo64': mo_part_64, \
+            'le_256_q': le_256_q, 're_256_q': re_256_q, 'mo_256_q': mo_256_q,\
+            'le_128_q': le_128_q, 're_128_q': re_128_q, 'mo_128_q': mo_128_q,\
+            'le_64_q': le_64_q, 're_64_q': re_64_q, 'mo_64_q': mo_64_q}
+
+
+class DMDNet(nn.Module):
+    def __init__(self, ngf = 64, banks_num = 128):
+        super().__init__()
+        self.part_sizes = np.array([80,80,50,110]) # size for 512
+        self.feature_sizes = np.array([256,128,64]) # size for 512
+
+        self.banks_num = banks_num
+        self.key_scale = 4
+
+        self.E_lq = FeatureExtractor(key_scale = self.key_scale)
+        self.E_hq = FeatureExtractor(key_scale = self.key_scale)
+
+        self.LE_256_KV = KeyValue(ngf, ngf // self.key_scale, ngf)
+        self.RE_256_KV = KeyValue(ngf, ngf // self.key_scale, ngf)
+        self.MO_256_KV = KeyValue(ngf, ngf // self.key_scale, ngf)
+
+        self.LE_128_KV = KeyValue(ngf * 2 , ngf * 2 // self.key_scale, ngf * 2)
+        self.RE_128_KV = KeyValue(ngf * 2 , ngf * 2 // self.key_scale, ngf * 2)
+        self.MO_128_KV = KeyValue(ngf * 2 , ngf * 2 // self.key_scale, ngf * 2)
+
+        self.LE_64_KV = KeyValue(ngf * 4 , ngf * 4 // self.key_scale, ngf * 4)
+        self.RE_64_KV = KeyValue(ngf * 4 , ngf * 4 // self.key_scale, ngf * 4)
+        self.MO_64_KV = KeyValue(ngf * 4 , ngf * 4 // self.key_scale, ngf * 4)
+
+
+        self.LE_256_Attention = AttentionBlock(64)
+        self.RE_256_Attention = AttentionBlock(64)
+        self.MO_256_Attention = AttentionBlock(64)
+
+        self.LE_128_Attention = AttentionBlock(128)
+        self.RE_128_Attention = AttentionBlock(128)
+        self.MO_128_Attention = AttentionBlock(128)
+
+        self.LE_64_Attention = AttentionBlock(256)
+        self.RE_64_Attention = AttentionBlock(256)
+        self.MO_64_Attention = AttentionBlock(256)
+
+        self.LE_256_Mask = MaskAttention(64)
+        self.RE_256_Mask = MaskAttention(64)
+        self.MO_256_Mask = MaskAttention(64)
+
+        self.LE_128_Mask = MaskAttention(128)
+        self.RE_128_Mask = MaskAttention(128)
+        self.MO_128_Mask = MaskAttention(128)
+
+        self.LE_64_Mask = MaskAttention(256)
+        self.RE_64_Mask = MaskAttention(256)
+        self.MO_64_Mask = MaskAttention(256)
+
+        self.MSDilate = MSDilateBlock(ngf*4, dilation = [4,3,2,1])
+
+        self.up1 = StyledUpBlock(ngf*4, ngf*2, noise_inject=False) #
+        self.up2 = StyledUpBlock(ngf*2, ngf, noise_inject=False) #
+        self.up3 = StyledUpBlock(ngf, ngf, noise_inject=False) #
+        self.up4 = nn.Sequential( 
+            SpectralNorm(nn.Conv2d(ngf, ngf, 3, 1, 1)),
+            nn.LeakyReLU(0.2),
+            UpResBlock(ngf),
+            UpResBlock(ngf),
+            SpectralNorm(nn.Conv2d(ngf, 3, kernel_size=3, stride=1, padding=1)),
+            nn.Tanh()
+        )
+ 
+        # define generic memory, revise register_buffer to register_parameter for backward update
+        self.register_buffer('le_256_mem_key', torch.randn(128,16,40,40))
+        self.register_buffer('re_256_mem_key', torch.randn(128,16,40,40))
+        self.register_buffer('mo_256_mem_key', torch.randn(128,16,55,55))
+        self.register_buffer('le_256_mem_value', torch.randn(128,64,40,40))
+        self.register_buffer('re_256_mem_value', torch.randn(128,64,40,40))
+        self.register_buffer('mo_256_mem_value', torch.randn(128,64,55,55))
+        
+
+        self.register_buffer('le_128_mem_key', torch.randn(128,32,20,20))
+        self.register_buffer('re_128_mem_key', torch.randn(128,32,20,20))
+        self.register_buffer('mo_128_mem_key', torch.randn(128,32,27,27))
+        self.register_buffer('le_128_mem_value', torch.randn(128,128,20,20))
+        self.register_buffer('re_128_mem_value', torch.randn(128,128,20,20))
+        self.register_buffer('mo_128_mem_value', torch.randn(128,128,27,27))
+
+        self.register_buffer('le_64_mem_key', torch.randn(128,64,10,10))
+        self.register_buffer('re_64_mem_key', torch.randn(128,64,10,10))
+        self.register_buffer('mo_64_mem_key', torch.randn(128,64,13,13))
+        self.register_buffer('le_64_mem_value', torch.randn(128,256,10,10))
+        self.register_buffer('re_64_mem_value', torch.randn(128,256,10,10))
+        self.register_buffer('mo_64_mem_value', torch.randn(128,256,13,13))
+
+    
+    def readMem(self, k, v, q):
+        sim = F.conv2d(q, k)
+        score = F.softmax(sim/sqrt(sim.size(1)), dim=1) #B * S * 1 * 1 6*128
+        sb,sn,sw,sh = score.size()
+        s_m = score.view(sb, -1).unsqueeze(1)#2*1*M
+        vb,vn,vw,vh = v.size()
+        v_in = v.view(vb, -1).repeat(sb,1,1)#2*M*(c*w*h)
+        mem_out = torch.bmm(s_m, v_in).squeeze(1).view(sb, vn, vw,vh)
+        max_inds = torch.argmax(score, dim=1).squeeze()
+        return mem_out, max_inds
+    
+
+    def memorize(self, img, locs):
+        fs = self.E_hq(img, locs)
+        LE256_key, LE256_value = self.LE_256_KV(fs['le256'])
+        RE256_key, RE256_value = self.RE_256_KV(fs['re256'])
+        MO256_key, MO256_value = self.MO_256_KV(fs['mo256'])
+
+        LE128_key, LE128_value = self.LE_128_KV(fs['le128'])
+        RE128_key, RE128_value = self.RE_128_KV(fs['re128'])
+        MO128_key, MO128_value = self.MO_128_KV(fs['mo128'])
+
+        LE64_key, LE64_value = self.LE_64_KV(fs['le64'])
+        RE64_key, RE64_value = self.RE_64_KV(fs['re64'])
+        MO64_key, MO64_value = self.MO_64_KV(fs['mo64'])
+
+        Mem256 = {'LE256Key': LE256_key, 'LE256Value': LE256_value, 'RE256Key': RE256_key, 'RE256Value': RE256_value,'MO256Key': MO256_key, 'MO256Value': MO256_value}
+        Mem128 = {'LE128Key': LE128_key, 'LE128Value': LE128_value, 'RE128Key': RE128_key, 'RE128Value': RE128_value,'MO128Key': MO128_key, 'MO128Value': MO128_value}
+        Mem64 = {'LE64Key': LE64_key, 'LE64Value': LE64_value, 'RE64Key': RE64_key, 'RE64Value': RE64_value,'MO64Key': MO64_key, 'MO64Value': MO64_value}
+ 
+        FS256 = {'LE256F':fs['le256'], 'RE256F':fs['re256'], 'MO256F':fs['mo256']}
+        FS128 = {'LE128F':fs['le128'], 'RE128F':fs['re128'], 'MO128F':fs['mo128']}
+        FS64 = {'LE64F':fs['le64'], 'RE64F':fs['re64'], 'MO64F':fs['mo64']}
+        
+        return Mem256, Mem128, Mem64
+
+    def enhancer(self, fs_in, sp_256=None, sp_128=None, sp_64=None):
+        le_256_q = fs_in['le_256_q']
+        re_256_q = fs_in['re_256_q']
+        mo_256_q = fs_in['mo_256_q']
+
+        le_128_q = fs_in['le_128_q']
+        re_128_q = fs_in['re_128_q']
+        mo_128_q = fs_in['mo_128_q']
+
+        le_64_q = fs_in['le_64_q']
+        re_64_q = fs_in['re_64_q']
+        mo_64_q = fs_in['mo_64_q']
+
+        
+        ####for 256
+        le_256_mem_g, le_256_inds = self.readMem(self.le_256_mem_key, self.le_256_mem_value, le_256_q)
+        re_256_mem_g, re_256_inds = self.readMem(self.re_256_mem_key, self.re_256_mem_value, re_256_q)
+        mo_256_mem_g, mo_256_inds = self.readMem(self.mo_256_mem_key, self.mo_256_mem_value, mo_256_q)
+
+        le_128_mem_g, le_128_inds = self.readMem(self.le_128_mem_key, self.le_128_mem_value, le_128_q)
+        re_128_mem_g, re_128_inds = self.readMem(self.re_128_mem_key, self.re_128_mem_value, re_128_q)
+        mo_128_mem_g, mo_128_inds = self.readMem(self.mo_128_mem_key, self.mo_128_mem_value, mo_128_q)
+
+        le_64_mem_g, le_64_inds = self.readMem(self.le_64_mem_key, self.le_64_mem_value, le_64_q)
+        re_64_mem_g, re_64_inds = self.readMem(self.re_64_mem_key, self.re_64_mem_value, re_64_q)
+        mo_64_mem_g, mo_64_inds = self.readMem(self.mo_64_mem_key, self.mo_64_mem_value, mo_64_q)
+
+        if sp_256 is not None and sp_128 is not None and sp_64 is not None:
+            le_256_mem_s, _ = self.readMem(sp_256['LE256Key'], sp_256['LE256Value'], le_256_q)
+            re_256_mem_s, _ = self.readMem(sp_256['RE256Key'], sp_256['RE256Value'], re_256_q)
+            mo_256_mem_s, _ = self.readMem(sp_256['MO256Key'], sp_256['MO256Value'], mo_256_q)
+            le_256_mask = self.LE_256_Mask(fs_in['le256'],le_256_mem_s,le_256_mem_g)
+            le_256_mem = le_256_mask*le_256_mem_s + (1-le_256_mask)*le_256_mem_g
+            re_256_mask = self.RE_256_Mask(fs_in['re256'],re_256_mem_s,re_256_mem_g)
+            re_256_mem = re_256_mask*re_256_mem_s + (1-re_256_mask)*re_256_mem_g
+            mo_256_mask = self.MO_256_Mask(fs_in['mo256'],mo_256_mem_s,mo_256_mem_g)
+            mo_256_mem = mo_256_mask*mo_256_mem_s + (1-mo_256_mask)*mo_256_mem_g
+
+            le_128_mem_s, _ = self.readMem(sp_128['LE128Key'], sp_128['LE128Value'], le_128_q)
+            re_128_mem_s, _ = self.readMem(sp_128['RE128Key'], sp_128['RE128Value'], re_128_q)
+            mo_128_mem_s, _ = self.readMem(sp_128['MO128Key'], sp_128['MO128Value'], mo_128_q)
+            le_128_mask = self.LE_128_Mask(fs_in['le128'],le_128_mem_s,le_128_mem_g)
+            le_128_mem = le_128_mask*le_128_mem_s + (1-le_128_mask)*le_128_mem_g
+            re_128_mask = self.RE_128_Mask(fs_in['re128'],re_128_mem_s,re_128_mem_g)
+            re_128_mem = re_128_mask*re_128_mem_s + (1-re_128_mask)*re_128_mem_g
+            mo_128_mask = self.MO_128_Mask(fs_in['mo128'],mo_128_mem_s,mo_128_mem_g)
+            mo_128_mem = mo_128_mask*mo_128_mem_s + (1-mo_128_mask)*mo_128_mem_g
+
+            le_64_mem_s, _ = self.readMem(sp_64['LE64Key'], sp_64['LE64Value'], le_64_q)
+            re_64_mem_s, _ = self.readMem(sp_64['RE64Key'], sp_64['RE64Value'], re_64_q)
+            mo_64_mem_s, _ = self.readMem(sp_64['MO64Key'], sp_64['MO64Value'], mo_64_q)
+            le_64_mask = self.LE_64_Mask(fs_in['le64'],le_64_mem_s,le_64_mem_g)
+            le_64_mem = le_64_mask*le_64_mem_s + (1-le_64_mask)*le_64_mem_g
+            re_64_mask = self.RE_64_Mask(fs_in['re64'],re_64_mem_s,re_64_mem_g)
+            re_64_mem = re_64_mask*re_64_mem_s + (1-re_64_mask)*re_64_mem_g
+            mo_64_mask = self.MO_64_Mask(fs_in['mo64'],mo_64_mem_s,mo_64_mem_g)
+            mo_64_mem = mo_64_mask*mo_64_mem_s + (1-mo_64_mask)*mo_64_mem_g
+        else:
+            le_256_mem = le_256_mem_g
+            re_256_mem = re_256_mem_g
+            mo_256_mem = mo_256_mem_g
+            le_128_mem = le_128_mem_g
+            re_128_mem = re_128_mem_g
+            mo_128_mem = mo_128_mem_g
+            le_64_mem = le_64_mem_g
+            re_64_mem = re_64_mem_g
+            mo_64_mem = mo_64_mem_g
+
+        le_256_mem_norm = adaptive_instance_normalization_4D(le_256_mem, fs_in['le256'])
+        re_256_mem_norm = adaptive_instance_normalization_4D(re_256_mem, fs_in['re256'])
+        mo_256_mem_norm = adaptive_instance_normalization_4D(mo_256_mem, fs_in['mo256'])
+        
+        ####for 128
+        le_128_mem_norm = adaptive_instance_normalization_4D(le_128_mem, fs_in['le128'])
+        re_128_mem_norm = adaptive_instance_normalization_4D(re_128_mem, fs_in['re128'])
+        mo_128_mem_norm = adaptive_instance_normalization_4D(mo_128_mem, fs_in['mo128'])
+        
+        ####for 64
+        le_64_mem_norm = adaptive_instance_normalization_4D(le_64_mem, fs_in['le64'])
+        re_64_mem_norm = adaptive_instance_normalization_4D(re_64_mem, fs_in['re64'])
+        mo_64_mem_norm = adaptive_instance_normalization_4D(mo_64_mem, fs_in['mo64'])
+    
+
+        EnMem256 = {'LE256Norm': le_256_mem_norm, 'RE256Norm': re_256_mem_norm, 'MO256Norm': mo_256_mem_norm}
+        EnMem128 = {'LE128Norm': le_128_mem_norm, 'RE128Norm': re_128_mem_norm, 'MO128Norm': mo_128_mem_norm}
+        EnMem64 = {'LE64Norm': le_64_mem_norm, 'RE64Norm': re_64_mem_norm, 'MO64Norm': mo_64_mem_norm}
+        Ind256 = {'LE': le_256_inds, 'RE': re_256_inds, 'MO': mo_256_inds}
+        Ind128 = {'LE': le_128_inds, 'RE': re_128_inds, 'MO': mo_128_inds}
+        Ind64 = {'LE': le_64_inds, 'RE': re_64_inds, 'MO': mo_64_inds}
+        return EnMem256, EnMem128, EnMem64, Ind256, Ind128, Ind64
+
+    def reconstruct(self, fs_in, locs, memstar):
+        le_256_mem_norm, re_256_mem_norm, mo_256_mem_norm = memstar[0]['LE256Norm'], memstar[0]['RE256Norm'], memstar[0]['MO256Norm']
+        le_128_mem_norm, re_128_mem_norm, mo_128_mem_norm = memstar[1]['LE128Norm'], memstar[1]['RE128Norm'], memstar[1]['MO128Norm']
+        le_64_mem_norm, re_64_mem_norm, mo_64_mem_norm = memstar[2]['LE64Norm'], memstar[2]['RE64Norm'], memstar[2]['MO64Norm']
+
+        le_256_final = self.LE_256_Attention(le_256_mem_norm - fs_in['le256']) * le_256_mem_norm + fs_in['le256']
+        re_256_final = self.RE_256_Attention(re_256_mem_norm - fs_in['re256']) * re_256_mem_norm + fs_in['re256']
+        mo_256_final = self.MO_256_Attention(mo_256_mem_norm - fs_in['mo256']) * mo_256_mem_norm + fs_in['mo256']
+        
+        le_128_final = self.LE_128_Attention(le_128_mem_norm - fs_in['le128']) * le_128_mem_norm + fs_in['le128']
+        re_128_final = self.RE_128_Attention(re_128_mem_norm - fs_in['re128']) * re_128_mem_norm + fs_in['re128']
+        mo_128_final = self.MO_128_Attention(mo_128_mem_norm - fs_in['mo128']) * mo_128_mem_norm + fs_in['mo128']
+        
+        le_64_final = self.LE_64_Attention(le_64_mem_norm - fs_in['le64']) * le_64_mem_norm + fs_in['le64']
+        re_64_final = self.RE_64_Attention(re_64_mem_norm - fs_in['re64']) * re_64_mem_norm + fs_in['re64']
+        mo_64_final = self.MO_64_Attention(mo_64_mem_norm - fs_in['mo64']) * mo_64_mem_norm + fs_in['mo64']
+
+
+        le_location = locs[:,0,:]
+        re_location = locs[:,1,:]
+        mo_location = locs[:,3,:]
+        le_location = le_location.cpu().int().numpy()
+        re_location = re_location.cpu().int().numpy()
+        mo_location = mo_location.cpu().int().numpy()
+
+        up_in_256 = fs_in['f256'].clone()# * 0
+        up_in_128 = fs_in['f128'].clone()# * 0
+        up_in_64 = fs_in['f64'].clone()# * 0
+
+        for i in range(fs_in['f256'].size(0)):
+            up_in_256[i:i+1,:,le_location[i,1]//2:le_location[i,3]//2,le_location[i,0]//2:le_location[i,2]//2] = F.interpolate(le_256_final[i:i+1,:,:,:].clone(), (le_location[i,3]//2-le_location[i,1]//2,le_location[i,2]//2-le_location[i,0]//2),mode='bilinear',align_corners=False)
+            up_in_256[i:i+1,:,re_location[i,1]//2:re_location[i,3]//2,re_location[i,0]//2:re_location[i,2]//2] = F.interpolate(re_256_final[i:i+1,:,:,:].clone(), (re_location[i,3]//2-re_location[i,1]//2,re_location[i,2]//2-re_location[i,0]//2),mode='bilinear',align_corners=False)
+            up_in_256[i:i+1,:,mo_location[i,1]//2:mo_location[i,3]//2,mo_location[i,0]//2:mo_location[i,2]//2] = F.interpolate(mo_256_final[i:i+1,:,:,:].clone(), (mo_location[i,3]//2-mo_location[i,1]//2,mo_location[i,2]//2-mo_location[i,0]//2),mode='bilinear',align_corners=False)
+            
+            up_in_128[i:i+1,:,le_location[i,1]//4:le_location[i,3]//4,le_location[i,0]//4:le_location[i,2]//4] = F.interpolate(le_128_final[i:i+1,:,:,:].clone(), (le_location[i,3]//4-le_location[i,1]//4,le_location[i,2]//4-le_location[i,0]//4),mode='bilinear',align_corners=False)
+            up_in_128[i:i+1,:,re_location[i,1]//4:re_location[i,3]//4,re_location[i,0]//4:re_location[i,2]//4] = F.interpolate(re_128_final[i:i+1,:,:,:].clone(), (re_location[i,3]//4-re_location[i,1]//4,re_location[i,2]//4-re_location[i,0]//4),mode='bilinear',align_corners=False)
+            up_in_128[i:i+1,:,mo_location[i,1]//4:mo_location[i,3]//4,mo_location[i,0]//4:mo_location[i,2]//4] = F.interpolate(mo_128_final[i:i+1,:,:,:].clone(), (mo_location[i,3]//4-mo_location[i,1]//4,mo_location[i,2]//4-mo_location[i,0]//4),mode='bilinear',align_corners=False)
+
+            up_in_64[i:i+1,:,le_location[i,1]//8:le_location[i,3]//8,le_location[i,0]//8:le_location[i,2]//8] = F.interpolate(le_64_final[i:i+1,:,:,:].clone(), (le_location[i,3]//8-le_location[i,1]//8,le_location[i,2]//8-le_location[i,0]//8),mode='bilinear',align_corners=False)
+            up_in_64[i:i+1,:,re_location[i,1]//8:re_location[i,3]//8,re_location[i,0]//8:re_location[i,2]//8] = F.interpolate(re_64_final[i:i+1,:,:,:].clone(), (re_location[i,3]//8-re_location[i,1]//8,re_location[i,2]//8-re_location[i,0]//8),mode='bilinear',align_corners=False)
+            up_in_64[i:i+1,:,mo_location[i,1]//8:mo_location[i,3]//8,mo_location[i,0]//8:mo_location[i,2]//8] = F.interpolate(mo_64_final[i:i+1,:,:,:].clone(), (mo_location[i,3]//8-mo_location[i,1]//8,mo_location[i,2]//8-mo_location[i,0]//8),mode='bilinear',align_corners=False)
+        
+        ms_in_64 = self.MSDilate(fs_in['f64'].clone())
+        fea_up1 = self.up1(ms_in_64, up_in_64)
+        fea_up2 = self.up2(fea_up1, up_in_128) #
+        fea_up3 = self.up3(fea_up2, up_in_256) #
+        output = self.up4(fea_up3) #
+        return output
+
+    def generate_specific_dictionary(self, sp_imgs=None, sp_locs=None):
+        return self.memorize(sp_imgs, sp_locs)
+
+    def forward(self, lq=None, loc=None, sp_256 = None, sp_128 = None, sp_64 = None):
+        fs_in = self.E_lq(lq, loc) # low quality images
+        GeMemNorm256, GeMemNorm128, GeMemNorm64, Ind256, Ind128, Ind64 = self.enhancer(fs_in)
+        GeOut = self.reconstruct(fs_in, loc, memstar = [GeMemNorm256, GeMemNorm128, GeMemNorm64])
+        if sp_256 is not None and sp_128 is not None and sp_64 is not None:
+            GSMemNorm256, GSMemNorm128, GSMemNorm64, _, _, _ = self.enhancer(fs_in, sp_256, sp_128, sp_64)
+            GSOut = self.reconstruct(fs_in, loc, memstar = [GSMemNorm256, GSMemNorm128, GSMemNorm64])
+        else:
+            GSOut = None
+        return GeOut, GSOut
+
+class UpResBlock(nn.Module):
+    def __init__(self, dim, conv_layer = nn.Conv2d, norm_layer = nn.BatchNorm2d):
+        super(UpResBlock, self).__init__()
+        self.Model = nn.Sequential(
+            SpectralNorm(conv_layer(dim, dim, 3, 1, 1)),
+            nn.LeakyReLU(0.2),
+            SpectralNorm(conv_layer(dim, dim, 3, 1, 1)),
+        )
+    def forward(self, x):
+        out = x + self.Model(x)
+        return out

roop/processors/Enhance_GFPGAN.py

@@ -0,0 +1,70 @@
+from typing import Any, List, Callable
+import cv2 
+import numpy as np
+import onnxruntime
+import roop.globals
+
+from roop.typing import Face, Frame
+from roop.utilities import resolve_relative_path
+
+
+# THREAD_LOCK = threading.Lock()
+
+
+class Enhance_GFPGAN():
+
+    model_gfpgan = None
+    name = None
+    devicename = None
+
+    processorname = 'gfpgan'
+    type = 'enhance'
+
+
+    def Initialize(self, devicename):
+        if self.model_gfpgan is None:
+            model_path = resolve_relative_path('../models/GFPGANv1.4.onnx')
+            self.model_gfpgan = onnxruntime.InferenceSession(model_path, None, providers=roop.globals.execution_providers)
+            self.devicename = devicename
+
+        self.name = self.model_gfpgan.get_inputs()[0].name
+
+    def Run(self, source_face: Face, target_face: Face, temp_frame: Frame) -> Frame:
+        # preprocess
+        input_size = temp_frame.shape[1]
+        temp_frame = cv2.resize(temp_frame, (512, 512), cv2.INTER_CUBIC)
+
+        temp_frame = cv2.cvtColor(temp_frame, cv2.COLOR_BGR2RGB)
+        temp_frame = temp_frame.astype('float32') / 255.0
+        temp_frame = (temp_frame - 0.5) / 0.5
+        temp_frame = np.expand_dims(temp_frame, axis=0).transpose(0, 3, 1, 2)
+
+        io_binding = self.model_gfpgan.io_binding()           
+        io_binding.bind_cpu_input("input", temp_frame)
+        io_binding.bind_output("1288", self.devicename)
+        self.model_gfpgan.run_with_iobinding(io_binding)
+        ort_outs = io_binding.copy_outputs_to_cpu()
+        result = ort_outs[0][0]
+
+        # post-process
+        result = np.clip(result, -1, 1)
+        result = (result + 1) / 2
+        result = result.transpose(1, 2, 0) * 255.0
+        result = cv2.cvtColor(result, cv2.COLOR_RGB2BGR)
+        scale_factor = int(result.shape[1] / input_size)       
+        return result.astype(np.uint8), scale_factor
+
+
+    def Release(self):
+        self.model_gfpgan = None
+
+
+
+
+
+
+
+
+
+
+

roop/processors/FaceSwapInsightFace.py

@@ -0,0 +1,40 @@
+from typing import Any, List, Callable
+import roop.globals
+import insightface
+import cv2
+import numpy as np
+
+from roop.typing import Face, Frame
+from roop.utilities import resolve_relative_path
+
+
+
+class FaceSwapInsightFace():
+    model_swap_insightface = None
+
+
+    processorname = 'faceswap'
+    type = 'swap'
+
+
+    def Initialize(self, devicename):
+        if self.model_swap_insightface is None:
+            model_path = resolve_relative_path('../models/inswapper_128.onnx')
+            self.model_swap_insightface = insightface.model_zoo.get_model(model_path, providers=roop.globals.execution_providers)
+
+    
+    def Run(self, source_face: Face, target_face: Face, temp_frame: Frame) -> Frame:
+        img_fake, M = self.model_swap_insightface.get(temp_frame, target_face, source_face, paste_back=False)
+        target_face.matrix = M
+        return img_fake 
+
+
+    def Release(self):
+        del self.model_swap_insightface
+        self.model_swap_insightface = None
+
+
+                
+
+
+

roop/processors/Mask_Clip2Seg.py

@@ -0,0 +1,93 @@
+import os
+import cv2
+import numpy as np
+import torch
+import threading
+from torchvision import transforms
+from clip.clipseg import CLIPDensePredT
+from numpy import asarray
+from typing import Any, List, Callable
+import numpy as np
+
+from roop.typing import Face, Frame
+from roop.utilities import resolve_relative_path
+
+THREAD_LOCK_CLIP = threading.Lock()
+
+
+class Mask_Clip2Seg():
+
+    model_clip = None
+
+    processorname = 'clip2seg'
+    type = 'mask'
+
+
+    def Initialize(self, devicename):
+        if self.model_clip is None:
+            self.model_clip = CLIPDensePredT(version='ViT-B/16', reduce_dim=64, complex_trans_conv=True)
+            self.model_clip.eval();
+            self.model_clip.load_state_dict(torch.load('models/CLIP/rd64-uni-refined.pth', map_location=torch.device('cpu')), strict=False)
+
+        device = torch.device(devicename)
+        self.model_clip.to(device)
+
+
+    def Run(self, img1, keywords:str) -> Frame:
+        if keywords is None or len(keywords) < 1:
+            return img1
+        
+        source_image_small = cv2.resize(img1, (256,256))
+        
+        img_mask = np.full((source_image_small.shape[0],source_image_small.shape[1]), 0, dtype=np.float32)
+        mask_border = 1
+        l = 0
+        t = 0
+        r = 1
+        b = 1
+        
+        mask_blur = 5
+        clip_blur = 5
+        
+        img_mask = cv2.rectangle(img_mask, (mask_border+int(l), mask_border+int(t)), 
+                                (256 - mask_border-int(r), 256-mask_border-int(b)), (255, 255, 255), -1)    
+        img_mask = cv2.GaussianBlur(img_mask, (mask_blur*2+1,mask_blur*2+1), 0)    
+        img_mask /= 255
+
+        
+        input_image = source_image_small
+
+        transform = transforms.Compose([
+            transforms.ToTensor(),
+            transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
+            transforms.Resize((256, 256)),
+        ])
+        img = transform(input_image).unsqueeze(0)
+
+        thresh = 0.5
+        prompts = keywords.split(',')
+        with THREAD_LOCK_CLIP:
+            with torch.no_grad():
+                preds = self.model_clip(img.repeat(len(prompts),1,1,1), prompts)[0]
+        clip_mask = torch.sigmoid(preds[0][0])
+        for i in range(len(prompts)-1):
+            clip_mask += torch.sigmoid(preds[i+1][0])
+           
+        clip_mask = clip_mask.data.cpu().numpy()
+        np.clip(clip_mask, 0, 1)
+        
+        clip_mask[clip_mask>thresh] = 1.0
+        clip_mask[clip_mask<=thresh] = 0.0
+        kernel = np.ones((5, 5), np.float32)
+        clip_mask = cv2.dilate(clip_mask, kernel, iterations=1)
+        clip_mask = cv2.GaussianBlur(clip_mask, (clip_blur*2+1,clip_blur*2+1), 0)
+       
+        img_mask *= clip_mask
+        img_mask[img_mask<0.0] = 0.0
+        return img_mask
+       
+
+
+    def Release(self):
+        self.model_clip = None
+

roop/processors/frame/face_swapper.py

@@ -0,0 +1,113 @@
+from typing import Any, List, Callable
+import cv2
+import insightface
+import threading
+
+import roop.globals
+import roop.processors.frame.core
+from roop.face_util import get_first_face, get_all_faces
+from roop.typing import Face, Frame
+from roop.utilities import conditional_download, resolve_relative_path, is_image, is_video, compute_cosine_distance, get_destfilename_from_path
+
+FACE_SWAPPER = None
+THREAD_LOCK = threading.Lock()
+NAME = 'ROOP.FACE-SWAPPER'
+
+DIST_THRESHOLD = 0.65
+
+
+def get_face_swapper() -> Any:
+    global FACE_SWAPPER
+
+    with THREAD_LOCK:
+        if FACE_SWAPPER is None:
+            model_path = resolve_relative_path('../models/inswapper_128.onnx')
+            FACE_SWAPPER = insightface.model_zoo.get_model(model_path, providers=roop.globals.execution_providers)
+    return FACE_SWAPPER
+
+
+def pre_check() -> bool:
+    download_directory_path = resolve_relative_path('../models')
+    conditional_download(download_directory_path, ['https://huggingface.co/countfloyd/deepfake/resolve/main/inswapper_128.onnx'])
+    return True
+
+
+def pre_start() -> bool:
+    return True
+
+
+def post_process() -> None:
+    global FACE_SWAPPER
+
+    FACE_SWAPPER = None
+
+
+def swap_face(source_face: Face, target_face: Face, temp_frame: Frame) -> Frame:
+    return get_face_swapper().get(temp_frame, target_face, source_face, paste_back=True)
+
+
+def process_frame(source_face: Face, target_face: Face, temp_frame: Frame) -> Frame:
+    global DIST_THRESHOLD
+
+    if roop.globals.many_faces:
+        many_faces = get_all_faces(temp_frame)
+        if many_faces is not None:
+            for target_face in many_faces:
+                if target_face['det_score'] > 0.65:
+                    temp_frame = swap_face(source_face, target_face, temp_frame)
+    else:
+        if target_face:
+            target_embedding = target_face.embedding
+            many_faces = get_all_faces(temp_frame)
+            target_face = None
+            for dest_face in many_faces:
+                dest_embedding = dest_face.embedding
+                if compute_cosine_distance(target_embedding, dest_embedding) <= DIST_THRESHOLD:
+                    target_face = dest_face
+                    break
+            if target_face:
+                temp_frame = swap_face(source_face, target_face, temp_frame)
+            return temp_frame
+                    
+        target_face = get_first_face(temp_frame)
+        if target_face is not None:
+            temp_frame = swap_face(source_face, target_face, temp_frame)
+    return temp_frame
+
+
+
+def process_frames(is_batch: bool, source_face: Face, target_face: Face, temp_frame_paths: List[str], update: Callable[[], None]) -> None:
+    for temp_frame_path in temp_frame_paths:
+        temp_frame = cv2.imread(temp_frame_path)
+        if temp_frame is not None:
+            result = process_frame(source_face, target_face, temp_frame)
+            if result is not None:
+                if is_batch:
+                    tf = get_destfilename_from_path(temp_frame_path, roop.globals.output_path, '_fake.png')
+                    cv2.imwrite(tf, result)
+                else:
+                    cv2.imwrite(temp_frame_path, result)
+        if update:
+            update()
+
+
+def process_image(source_face: Any, target_face: Any, target_path: str, output_path: str) -> None:
+    global DIST_THRESHOLD
+
+    target_frame = cv2.imread(target_path)
+    if target_frame is not None:
+        result = process_frame(source_face, target_face, target_frame)
+        if result is not None:
+            cv2.imwrite(output_path, result)
+
+
+def process_video(source_face: Any, target_face: Any, temp_frame_paths: List[str]) -> None:
+    global DIST_THRESHOLD
+
+    roop.processors.frame.core.process_video(source_face, target_face, temp_frame_paths, process_frames)
+
+
+def process_batch_images(source_face: Any, target_face: Any, temp_frame_paths: List[str]) -> None:
+    global DIST_THRESHOLD
+
+    roop.processors.frame.core.process_batch(source_face, target_face, temp_frame_paths, process_frames)

roop/template_parser.py

@@ -0,0 +1,23 @@
+import re
+from datetime import datetime
+
+template_functions = {
+    "timestamp": lambda data: str(int(datetime.now().timestamp())),
+    "i": lambda data: data.get("index", False),
+    "file": lambda data: data.get("file", False),
+    "date": lambda data: datetime.now().strftime("%Y-%m-%d"),
+    "time": lambda data: datetime.now().strftime("%H-%M-%S"),
+}
+
+
+def parse(text: str, data: dict):
+    pattern = r"\{([^}]+)\}"
+
+    matches = re.findall(pattern, text)
+
+    for match in matches:
+        replacement = template_functions[match](data)
+        if replacement is not False:
+            text = text.replace(f"{{{match}}}", replacement)
+
+    return text

roop/typing.py

@@ -0,0 +1,7 @@
+from typing import Any
+
+from insightface.app.common import Face
+import numpy
+
+Face = Face
+Frame = numpy.ndarray[Any, Any]

roop/ui.py

@@ -0,0 +1,926 @@
+import os
+import time
+import gradio as gr
+import cv2
+import pathlib
+import shutil
+import roop.globals
+import roop.metadata
+import roop.utilities as util
+
+from roop.face_util import extract_face_images
+from roop.capturer import get_video_frame, get_video_frame_total, get_image_frame
+from roop.ProcessEntry import ProcessEntry
+
+restart_server = False
+live_cam_active = False
+
+RECENT_DIRECTORY_SOURCE = None
+RECENT_DIRECTORY_TARGET = None
+RECENT_DIRECTORY_OUTPUT = None
+
+SELECTION_FACES_DATA = None
+
+last_image = None
+
+input_thumbs = []
+target_thumbs = []
+
+
+IS_INPUT = True
+SELECTED_FACE_INDEX = 0
+
+SELECTED_INPUT_FACE_INDEX = 0
+SELECTED_TARGET_FACE_INDEX = 0
+
+roop.globals.keep_fps = None
+roop.globals.keep_frames = None
+roop.globals.skip_audio = None
+roop.globals.use_batch = None
+
+input_faces = None
+target_faces = None
+face_selection = None
+fake_cam_image = None
+
+current_cam_image = None
+cam_swapping = False
+camthread = None
+
+selected_preview_index = 0
+
+is_processing = False            
+
+list_files_process : list[ProcessEntry] = []
+
+
+def prepare_environment():
+    roop.globals.output_path = os.path.abspath(os.path.join(os.getcwd(), "output"))
+    os.makedirs(roop.globals.output_path, exist_ok=True)
+    os.environ["TEMP"] = os.environ["TMP"] = os.path.abspath(os.path.join(os.getcwd(), "temp"))
+    os.makedirs(os.environ["TEMP"], exist_ok=True)
+    os.environ["GRADIO_TEMP_DIR"] = os.environ["TEMP"]
+
+
+def run():
+    from roop.core import suggest_execution_providers, decode_execution_providers, set_display_ui
+    global input_faces, target_faces, face_selection, fake_cam_image, restart_server, live_cam_active, on_settings_changed
+
+    prepare_environment()
+
+    available_themes = ["Default", "gradio/glass", "gradio/monochrome", "gradio/seafoam", "gradio/soft", "gstaff/xkcd", "freddyaboulton/dracula_revamped", "ysharma/steampunk"]
+    image_formats = ['jpg','png', 'webp']
+    video_formats = ['avi','mkv', 'mp4', 'webm']
+    video_codecs = ['libx264', 'libx265', 'libvpx-vp9', 'h264_nvenc', 'hevc_nvenc']
+    providerlist = suggest_execution_providers()
+    
+    settings_controls = []
+
+    live_cam_active = roop.globals.CFG.live_cam_start_active
+    set_display_ui(show_msg)
+    roop.globals.execution_providers = decode_execution_providers([roop.globals.CFG.provider])
+    print(f'Using provider {roop.globals.execution_providers} - Device:{util.get_device()}')    
+    
+    run_server = True
+    mycss = """
+        span {color: var(--block-info-text-color)}
+        #fixedheight {
+            max-height: 238.4px;
+            overflow-y: auto !important;
+        }
+"""
+
+    while run_server:
+        server_name = roop.globals.CFG.server_name
+        if server_name is None or len(server_name) < 1:
+            server_name = None
+        server_port = roop.globals.CFG.server_port
+        if server_port <= 0:
+            server_port = None
+        ssl_verify = False if server_name == '0.0.0.0' else True
+        with gr.Blocks(title=f'{roop.metadata.name} {roop.metadata.version}', theme=roop.globals.CFG.selected_theme, css=mycss) as ui:
+            with gr.Row(variant='compact'):
+                    gr.Markdown(f"### [{roop.metadata.name} {roop.metadata.version}](https://github.com/C0untFloyd/roop-unleashed)")
+                    gr.HTML(util.create_version_html(), elem_id="versions")
+            with gr.Tab("🎭 Face Swap"):
+                with gr.Row(variant='panel'):
+                    with gr.Column(scale=2):
+                        with gr.Row():
+                            with gr.Column(min_width=160):
+                                input_faces = gr.Gallery(label="Input faces", allow_preview=True, preview=True, height=128, object_fit="scale-down")
+                                mask_top = gr.Slider(0, 256, value=0, label="Offset Face Top", step=1.0, interactive=True)
+                                bt_remove_selected_input_face = gr.Button("❌ Remove selected", size='sm')
+                                bt_clear_input_faces = gr.Button("💥 Clear all", variant='stop', size='sm')
+                            with gr.Column(min_width=160):
+                                target_faces = gr.Gallery(label="Target faces", allow_preview=True, preview=True, height=128, object_fit="scale-down")
+                                bt_remove_selected_target_face = gr.Button("❌ Remove selected", size='sm')
+                                bt_add_local = gr.Button('Add local files from', size='sm')
+                                local_folder = gr.Textbox(show_label=False, placeholder="/content/", interactive=True)
+                        with gr.Row(variant='panel'):
+                            bt_srcimg = gr.Image(label='Source Face Image', type='filepath', tool=None, height=233)
+                            bt_destfiles = gr.Files(label='Target File(s)', file_count="multiple", elem_id='filelist', height=233)
+                        with gr.Row(variant='panel'):
+                            gr.Markdown('')
+                            forced_fps = gr.Slider(minimum=0, maximum=120, value=0, label="Video FPS", info='Overrides detected fps if not 0', step=1.0, interactive=True, container=True)
+    
+                    with gr.Column(scale=2):
+                        previewimage = gr.Image(label="Preview Image", height=576, interactive=False)
+                        with gr.Row(variant='panel'):
+                                fake_preview = gr.Checkbox(label="Face swap frames", value=False)
+                                bt_refresh_preview = gr.Button("🔄 Refresh", variant='secondary', size='sm')
+                                bt_use_face_from_preview = gr.Button("Use Face from this Frame", variant='primary', size='sm')
+                        with gr.Row():
+                            preview_frame_num = gr.Slider(0, 0, value=0, label="Frame Number", step=1.0, interactive=True)
+                        with gr.Row():
+                            text_frame_clip = gr.Markdown('Processing frame range [0 - 0]')
+                            set_frame_start = gr.Button("⬅ Set as Start", size='sm')
+                            set_frame_end = gr.Button("➡ Set as End", size='sm')
+                with gr.Row(visible=False) as dynamic_face_selection:
+                    with gr.Column(scale=2):
+                        face_selection = gr.Gallery(label="Detected faces", allow_preview=True, preview=True, height=256, object_fit="scale-down")
+                    with gr.Column():
+                        bt_faceselect = gr.Button("☑ Use selected face", size='sm')
+                        bt_cancelfaceselect = gr.Button("Done", size='sm')
+                    with gr.Column():
+                        gr.Markdown(' ') 
+            
+                with gr.Row(variant='panel'):
+                    with gr.Column(scale=1):
+                        selected_face_detection = gr.Dropdown(["First found", "All faces", "Selected face", "All female", "All male"], value="First found", label="Select face selection for swapping")
+                        max_face_distance = gr.Slider(0.01, 1.0, value=0.65, label="Max Face Similarity Threshold")
+                        video_swapping_method = gr.Dropdown(["Extract Frames to media","In-Memory processing"], value="In-Memory", label="Select video processing method", interactive=True)
+                        roop.globals.keep_frames = gr.Checkbox(label="Keep Frames (relevant only when extracting frames)", value=False)
+                        roop.globals.skip_audio = gr.Checkbox(label="Skip audio", value=False)
+                    with gr.Column(scale=1):
+                        selected_enhancer = gr.Dropdown(["None", "Codeformer", "DMDNet", "GFPGAN"], value="None", label="Select post-processing")
+                        blend_ratio = gr.Slider(0.0, 1.0, value=0.65, label="Original/Enhanced image blend ratio")
+                    with gr.Column(scale=1):
+                        chk_useclip = gr.Checkbox(label="Use Text Masking", value=False)
+                        clip_text = gr.Textbox(label="List of objects to mask and restore back on fake image", placeholder="cup,hands,hair,banana" ,elem_id='tooltip')
+                        gr.Dropdown(["Clip2Seg"], value="Clip2Seg", label="Engine")
+                        bt_preview_mask = gr.Button("👥 Show Mask Preview", variant='secondary')
+                            
+                with gr.Row(variant='panel'):
+                    with gr.Column():
+                        bt_start = gr.Button("▶ Start", variant='primary')
+                        gr.Button("👀 Open Output Folder", size='sm').click(fn=lambda: util.open_folder(roop.globals.output_path))
+                    with gr.Column():
+                        bt_stop = gr.Button("⏹ Stop", variant='secondary')
+                    with gr.Column(scale=2):
+                        gr.Markdown(' ') 
+                with gr.Row(variant='panel'):
+                    with gr.Column():
+                        resultfiles = gr.Files(label='Processed File(s)', interactive=False)
+                    with gr.Column():
+                        resultimage = gr.Image(type='filepath', label='Final Image', interactive=False, )
+                                
+                        
+            with gr.Tab("🎥 Live Cam"):
+                with gr.Row():
+                    with gr.Column(scale=2):
+                        cam_toggle = gr.Checkbox(label='Activate', value=live_cam_active)
+                    with gr.Column(scale=1):
+                        vcam_toggle = gr.Checkbox(label='Stream to virtual camera', value=False)
+                    with gr.Column(scale=1):
+                        camera_num = gr.Slider(0, 2, value=0, label="Camera Number", step=1.0, interactive=True)                       
+
+                if live_cam_active:
+                    with gr.Row():
+                        with gr.Column():
+                            cam = gr.Webcam(label='Camera', source='webcam', mirror_webcam=True, interactive=True, streaming=False)
+                        with gr.Column():
+                            fake_cam_image = gr.Image(label='Fake Camera Output', interactive=False)
+
+
+            with gr.Tab("🎉 Extras"):
+                with gr.Row():
+                    files_to_process = gr.Files(label='File(s) to process', file_count="multiple")
+                # with gr.Row(variant='panel'):
+                #     with gr.Accordion(label="Post process", open=False):
+                #         with gr.Column():
+                #             selected_post_enhancer = gr.Dropdown(["None", "Codeformer", "GFPGAN"], value="None", label="Select post-processing")
+                #         with gr.Column():
+                #             gr.Button("Start").click(fn=lambda: gr.Info('Not yet implemented...'))
+                with gr.Row(variant='panel'):
+                    with gr.Accordion(label="Video/GIF", open=False):
+                        with gr.Row(variant='panel'):
+                            with gr.Column():
+                                gr.Markdown("""
+                                            # Cut video
+                                            Be aware that this means re-encoding the video which might take a longer time.
+                                            Encoding uses your configuration from the Settings Tab.
+    """)
+                            with gr.Column():
+                                cut_start_time = gr.Slider(0, 1000000, value=0, label="Start Frame", step=1.0, interactive=True)
+                            with gr.Column():
+                                cut_end_time = gr.Slider(1, 1000000, value=1, label="End Frame", step=1.0, interactive=True)
+                            with gr.Column():
+                                start_cut_video = gr.Button("Start")
+
+    #                     with gr.Row(variant='panel'):
+    #                         with gr.Column():
+    #                             gr.Markdown("""
+    #                                         # Join videos
+    #                                         This also re-encodes the videos like cutting above.
+    # """)
+    #                         with gr.Column():
+    #                             start_join_videos = gr.Button("Start")
+                        with gr.Row(variant='panel'):
+                            gr.Markdown("Extract frames from video")
+                            start_extract_frames = gr.Button("Start")
+                        with gr.Row(variant='panel'):
+                            gr.Markdown("Create video from image files")
+                            gr.Button("Start").click(fn=lambda: gr.Info('Not yet implemented...'))
+                        with gr.Row(variant='panel'):
+                            gr.Markdown("Create GIF from video")
+                            start_create_gif = gr.Button("Create GIF")
+                with gr.Row():
+                    extra_files_output = gr.Files(label='Resulting output files', file_count="multiple")
+                        
+            
+            with gr.Tab("⚙ Settings"):
+                with gr.Row():
+                    with gr.Column():
+                        themes = gr.Dropdown(available_themes, label="Theme", info="Change needs complete restart", value=roop.globals.CFG.selected_theme)
+                    with gr.Column():
+                        settings_controls.append(gr.Checkbox(label="Public Server", value=roop.globals.CFG.server_share, elem_id='server_share', interactive=True))
+                        settings_controls.append(gr.Checkbox(label='Clear output folder before each run', value=roop.globals.CFG.clear_output, elem_id='clear_output', interactive=True))
+                        output_template = gr.Textbox(label="Filename Output Template", info="(file extension is added automatically)", lines=1, placeholder='{file}_{time}', value=roop.globals.CFG.output_template)
+                    with gr.Column():
+                        input_server_name = gr.Textbox(label="Server Name", lines=1, info="Leave blank to run locally", value=roop.globals.CFG.server_name)
+                    with gr.Column():
+                        input_server_port = gr.Number(label="Server Port", precision=0, info="Leave at 0 to use default", value=roop.globals.CFG.server_port)
+                with gr.Row():
+                    with gr.Column():
+                        settings_controls.append(gr.Dropdown(providerlist, label="Provider", value=roop.globals.CFG.provider, elem_id='provider', interactive=True))
+                        chk_det_size = gr.Checkbox(label="Use default Det-Size", value=True, elem_id='default_det_size', interactive=True)
+                        settings_controls.append(gr.Checkbox(label="Force CPU for Face Analyser", value=roop.globals.CFG.force_cpu, elem_id='force_cpu', interactive=True))
+                        max_threads = gr.Slider(1, 32, value=roop.globals.CFG.max_threads, label="Max. Number of Threads", info='default: 3', step=1.0, interactive=True)
+                    with gr.Column():
+                        memory_limit = gr.Slider(0, 128, value=roop.globals.CFG.memory_limit, label="Max. Memory to use (Gb)", info='0 meaning no limit', step=1.0, interactive=True)
+                        settings_controls.append(gr.Dropdown(image_formats, label="Image Output Format", info='default: png', value=roop.globals.CFG.output_image_format, elem_id='output_image_format', interactive=True))
+                    with gr.Column():
+                        settings_controls.append(gr.Dropdown(video_codecs, label="Video Codec", info='default: libx264', value=roop.globals.CFG.output_video_codec, elem_id='output_video_codec', interactive=True))
+                        settings_controls.append(gr.Dropdown(video_formats, label="Video Output Format", info='default: mp4', value=roop.globals.CFG.output_video_format, elem_id='output_video_format', interactive=True))
+                        video_quality = gr.Slider(0, 100, value=roop.globals.CFG.video_quality, label="Video Quality (crf)", info='default: 14', step=1.0, interactive=True)
+                    with gr.Column():
+                        button_apply_restart = gr.Button("Restart Server", variant='primary')
+                        settings_controls.append(gr.Checkbox(label='Start with active live cam', value=roop.globals.CFG.live_cam_start_active, elem_id='live_cam_start_active', interactive=True))
+                        button_clean_temp = gr.Button("Clean temp folder")
+                        button_apply_settings = gr.Button("Apply Settings")
+
+            previewinputs = [preview_frame_num, bt_destfiles, fake_preview, selected_enhancer, selected_face_detection,
+                                max_face_distance, blend_ratio, chk_useclip, clip_text] 
+            input_faces.select(on_select_input_face, None, None).then(fn=on_preview_frame_changed, inputs=previewinputs, outputs=[previewimage, mask_top])
+            bt_remove_selected_input_face.click(fn=remove_selected_input_face, outputs=[input_faces])
+            bt_srcimg.change(fn=on_srcimg_changed, show_progress='full', inputs=bt_srcimg, outputs=[dynamic_face_selection, face_selection, input_faces])
+
+            mask_top.input(fn=on_mask_top_changed, inputs=[mask_top], show_progress='hidden')
+
+
+            target_faces.select(on_select_target_face, None, None)
+            bt_remove_selected_target_face.click(fn=remove_selected_target_face, outputs=[target_faces])
+
+            forced_fps.change(fn=on_fps_changed, inputs=[forced_fps], show_progress='hidden')
+            bt_destfiles.change(fn=on_destfiles_changed, inputs=[bt_destfiles], outputs=[preview_frame_num, text_frame_clip], show_progress='hidden').then(fn=on_preview_frame_changed, inputs=previewinputs, outputs=[previewimage, mask_top], show_progress='full')
+            bt_destfiles.select(fn=on_destfiles_selected, outputs=[preview_frame_num, text_frame_clip, forced_fps], show_progress='hidden').then(fn=on_preview_frame_changed, inputs=previewinputs, outputs=[previewimage, mask_top], show_progress='hidden')
+            bt_destfiles.clear(fn=on_clear_destfiles, outputs=[target_faces])
+            resultfiles.select(fn=on_resultfiles_selected, inputs=[resultfiles], outputs=[resultimage])
+
+            face_selection.select(on_select_face, None, None)
+            bt_faceselect.click(fn=on_selected_face, outputs=[input_faces, target_faces, selected_face_detection])
+            bt_cancelfaceselect.click(fn=on_end_face_selection, outputs=[dynamic_face_selection, face_selection])
+            
+            bt_clear_input_faces.click(fn=on_clear_input_faces, outputs=[input_faces])
+
+            chk_det_size.select(fn=on_option_changed)
+
+            bt_add_local.click(fn=on_add_local_folder, inputs=[local_folder], outputs=[bt_destfiles])
+            bt_preview_mask.click(fn=on_preview_mask, inputs=[preview_frame_num, bt_destfiles, clip_text], outputs=[previewimage]) 
+
+            start_event = bt_start.click(fn=start_swap, 
+                inputs=[selected_enhancer, selected_face_detection, roop.globals.keep_frames,
+                         roop.globals.skip_audio, max_face_distance, blend_ratio, chk_useclip, clip_text,video_swapping_method],
+                outputs=[bt_start, resultfiles]).then(fn=on_resultfiles_finished, inputs=[resultfiles], outputs=[resultimage])
+            
+            bt_stop.click(fn=stop_swap, cancels=[start_event])
+            
+            bt_refresh_preview.click(fn=on_preview_frame_changed, inputs=previewinputs, outputs=[previewimage, mask_top])            
+            fake_preview.change(fn=on_preview_frame_changed, inputs=previewinputs, outputs=[previewimage, mask_top])
+            preview_frame_num.change(fn=on_preview_frame_changed, inputs=previewinputs, outputs=[previewimage, mask_top], show_progress='hidden')
+            bt_use_face_from_preview.click(fn=on_use_face_from_selected, show_progress='full', inputs=[bt_destfiles, preview_frame_num], outputs=[dynamic_face_selection, face_selection, target_faces, selected_face_detection])
+            set_frame_start.click(fn=on_set_frame, inputs=[set_frame_start, preview_frame_num], outputs=[text_frame_clip])
+            set_frame_end.click(fn=on_set_frame, inputs=[set_frame_end, preview_frame_num], outputs=[text_frame_clip])
+            
+            
+            # Live Cam
+            cam_toggle.change(fn=on_cam_toggle, inputs=[cam_toggle])
+
+            if live_cam_active:
+                vcam_toggle.change(fn=on_vcam_toggle, inputs=[vcam_toggle, camera_num], outputs=[cam, fake_cam_image])
+                cam.stream(on_stream_swap_cam, inputs=[cam, selected_enhancer, blend_ratio], outputs=[fake_cam_image], preprocess=True, postprocess=True, show_progress="hidden")
+
+            # Extras
+            start_cut_video.click(fn=on_cut_video, inputs=[files_to_process, cut_start_time, cut_end_time], outputs=[extra_files_output])
+            # start_join_videos.click(fn=on_join_videos, inputs=[files_to_process], outputs=[extra_files_output])
+            start_extract_frames.click(fn=on_extract_frames, inputs=[files_to_process], outputs=[extra_files_output])
+            start_create_gif.click(fn=on_create_gif, inputs=[files_to_process], outputs=[extra_files_output])
+
+            # Settings
+            for s in settings_controls:
+                s.select(fn=on_settings_changed)
+            max_threads.input(fn=lambda a,b='max_threads':on_settings_changed_misc(a,b), inputs=[max_threads])
+            memory_limit.input(fn=lambda a,b='memory_limit':on_settings_changed_misc(a,b), inputs=[memory_limit])
+            video_quality.input(fn=lambda a,b='video_quality':on_settings_changed_misc(a,b), inputs=[video_quality])
+
+            button_clean_temp.click(fn=clean_temp, outputs=[bt_srcimg, input_faces, target_faces, bt_destfiles])
+            button_apply_settings.click(apply_settings, inputs=[themes, input_server_name, input_server_port, output_template])
+            button_apply_restart.click(restart)
+
+
+
+        restart_server = False
+        try:
+            ui.queue().launch(inbrowser=True, server_name=server_name, server_port=server_port, share=roop.globals.CFG.server_share, ssl_verify=ssl_verify, prevent_thread_lock=True, show_error=True)
+        except:
+            restart_server = True
+            run_server = False
+        try:
+            while restart_server == False:
+                time.sleep(1.0)
+        except (KeyboardInterrupt, OSError):
+            print("Keyboard interruption in main thread... closing server.")
+            run_server = False
+        ui.close()
+
+
+def on_mask_top_changed(mask_top):
+    global SELECTED_INPUT_FACE_INDEX
+
+    if len(roop.globals.INPUT_FACES) > SELECTED_INPUT_FACE_INDEX:
+        roop.globals.INPUT_FACES[SELECTED_INPUT_FACE_INDEX].mask_top = mask_top
+
+
+def on_option_changed(evt: gr.SelectData):
+    attribname = evt.target.elem_id
+    if isinstance(evt.target, gr.Checkbox):
+        if hasattr(roop.globals, attribname):
+            setattr(roop.globals, attribname, evt.selected)
+            return
+    elif isinstance(evt.target, gr.Dropdown):
+        if hasattr(roop.globals, attribname):
+            setattr(roop.globals, attribname, evt.value)
+            return
+    raise gr.Error(f'Unhandled Setting for {evt.target}')
+
+
+def on_settings_changed_misc(new_val, attribname):
+    if hasattr(roop.globals.CFG, attribname):
+        setattr(roop.globals.CFG, attribname, new_val)
+    else:
+        print("Didn't find attrib!")
+        
+
+
+def on_settings_changed(evt: gr.SelectData):
+    attribname = evt.target.elem_id
+    if isinstance(evt.target, gr.Checkbox):
+        if hasattr(roop.globals.CFG, attribname):
+            setattr(roop.globals.CFG, attribname, evt.selected)
+            return
+    elif isinstance(evt.target, gr.Dropdown):
+        if hasattr(roop.globals.CFG, attribname):
+            setattr(roop.globals.CFG, attribname, evt.value)
+            return
+            
+    raise gr.Error(f'Unhandled Setting for {evt.target}')
+
+
+def on_add_local_folder(folder):
+    files = util.get_local_files_from_folder(folder)
+    if files is None:
+        gr.Warning("Empty folder or folder not found!")
+    return files
+
+
+def on_srcimg_changed(imgsrc, progress=gr.Progress()):
+    global RECENT_DIRECTORY_SOURCE, SELECTION_FACES_DATA, IS_INPUT, input_faces, face_selection, input_thumbs, last_image
+    
+    IS_INPUT = True
+
+    if imgsrc == None or last_image == imgsrc:
+        return gr.Column.update(visible=False), None, input_thumbs
+    
+    last_image = imgsrc
+    
+    progress(0, desc="Retrieving faces from image", )      
+    source_path = imgsrc
+    thumbs = []
+    if util.is_image(source_path):
+        roop.globals.source_path = source_path
+        RECENT_DIRECTORY_SOURCE = os.path.dirname(roop.globals.source_path)
+        SELECTION_FACES_DATA = extract_face_images(roop.globals.source_path,  (False, 0))
+        progress(0.5, desc="Retrieving faces from image")      
+        for f in SELECTION_FACES_DATA:
+            image = convert_to_gradio(f[1])
+            thumbs.append(image)
+            
+    progress(1.0, desc="Retrieving faces from image")      
+    if len(thumbs) < 1:
+        raise gr.Error('No faces detected!')
+
+    if len(thumbs) == 1:
+        face = SELECTION_FACES_DATA[0][0]
+        face.mask_top = 0
+        roop.globals.INPUT_FACES.append(face)
+        input_thumbs.append(thumbs[0])
+        return gr.Column.update(visible=False), None, input_thumbs
+       
+    return gr.Column.update(visible=True), thumbs, gr.Gallery.update(visible=True)
+
+def on_select_input_face(evt: gr.SelectData):
+    global SELECTED_INPUT_FACE_INDEX
+
+    SELECTED_INPUT_FACE_INDEX = evt.index
+
+
+def remove_selected_input_face():
+    global input_thumbs, SELECTED_INPUT_FACE_INDEX
+
+    if len(roop.globals.INPUT_FACES) > SELECTED_INPUT_FACE_INDEX:
+        f = roop.globals.INPUT_FACES.pop(SELECTED_INPUT_FACE_INDEX)
+        del f
+    if len(input_thumbs) > SELECTED_INPUT_FACE_INDEX:
+        f = input_thumbs.pop(SELECTED_INPUT_FACE_INDEX)
+        del f
+
+    return input_thumbs
+
+def on_select_target_face(evt: gr.SelectData):
+    global SELECTED_TARGET_FACE_INDEX
+
+    SELECTED_TARGET_FACE_INDEX = evt.index
+
+def remove_selected_target_face():
+    global target_thumbs, SELECTED_TARGET_FACE_INDEX
+
+    if len(roop.globals.TARGET_FACES) > SELECTED_TARGET_FACE_INDEX:
+        f = roop.globals.TARGET_FACES.pop(SELECTED_TARGET_FACE_INDEX)
+        del f
+    if len(target_thumbs) > SELECTED_TARGET_FACE_INDEX:
+        f = target_thumbs.pop(SELECTED_TARGET_FACE_INDEX)
+        del f
+    return target_thumbs
+
+
+
+
+
+def on_use_face_from_selected(files, frame_num):
+    global IS_INPUT, SELECTION_FACES_DATA
+
+    IS_INPUT = False
+    thumbs = []
+    
+    roop.globals.target_path = files[selected_preview_index].name
+    if util.is_image(roop.globals.target_path) and not roop.globals.target_path.lower().endswith(('gif')):
+        SELECTION_FACES_DATA = extract_face_images(roop.globals.target_path,  (False, 0))
+        if len(SELECTION_FACES_DATA) > 0:
+            for f in SELECTION_FACES_DATA:
+                image = convert_to_gradio(f[1])
+                thumbs.append(image)
+        else:
+            gr.Info('No faces detected!')
+            roop.globals.target_path = None
+                
+    elif util.is_video(roop.globals.target_path) or roop.globals.target_path.lower().endswith(('gif')):
+        selected_frame = frame_num
+        SELECTION_FACES_DATA = extract_face_images(roop.globals.target_path, (True, selected_frame))
+        if len(SELECTION_FACES_DATA) > 0:
+            for f in SELECTION_FACES_DATA:
+                image = convert_to_gradio(f[1])
+                thumbs.append(image)
+        else:
+            gr.Info('No faces detected!')
+            roop.globals.target_path = None
+
+    if len(thumbs) == 1:
+        roop.globals.TARGET_FACES.append(SELECTION_FACES_DATA[0][0])
+        target_thumbs.append(thumbs[0])
+        return gr.Row.update(visible=False), None, target_thumbs, gr.Dropdown.update(value='Selected face')
+
+    return gr.Row.update(visible=True), thumbs, gr.Gallery.update(visible=True), gr.Dropdown.update(visible=True)
+
+
+
+def on_select_face(evt: gr.SelectData):  # SelectData is a subclass of EventData
+    global SELECTED_FACE_INDEX
+    SELECTED_FACE_INDEX = evt.index
+    
+
+def on_selected_face():
+    global IS_INPUT, SELECTED_FACE_INDEX, SELECTION_FACES_DATA, input_thumbs, target_thumbs
+    
+    fd = SELECTION_FACES_DATA[SELECTED_FACE_INDEX]
+    image = convert_to_gradio(fd[1])
+    if IS_INPUT:
+        fd[0].mask_top = 0
+        roop.globals.INPUT_FACES.append(fd[0])
+        input_thumbs.append(image)
+        return input_thumbs, gr.Gallery.update(visible=True), gr.Dropdown.update(visible=True)
+    else:
+        roop.globals.TARGET_FACES.append(fd[0])
+        target_thumbs.append(image)
+        return gr.Gallery.update(visible=True), target_thumbs, gr.Dropdown.update(value='Selected face')
+    
+#        bt_faceselect.click(fn=on_selected_face, outputs=[dynamic_face_selection, face_selection, input_faces, target_faces])
+
+def on_end_face_selection():
+    return gr.Column.update(visible=False), None
+
+
+def on_preview_frame_changed(frame_num, files, fake_preview, enhancer, detection, face_distance, blend_ratio, use_clip, clip_text):
+    global SELECTED_INPUT_FACE_INDEX, is_processing
+
+    from roop.core import live_swap
+
+    mask_top = 0
+    if len(roop.globals.INPUT_FACES) > SELECTED_INPUT_FACE_INDEX:
+        if hasattr(roop.globals.INPUT_FACES[SELECTED_INPUT_FACE_INDEX], 'mask_top'):
+            mask_top = roop.globals.INPUT_FACES[SELECTED_INPUT_FACE_INDEX].mask_top
+        else:
+            roop.globals.INPUT_FACES[SELECTED_INPUT_FACE_INDEX].mask_top = mask_top
+
+    if is_processing or files is None or selected_preview_index >= len(files) or frame_num is None:
+        return None, mask_top
+
+    filename = files[selected_preview_index].name
+    if util.is_video(filename) or filename.lower().endswith('gif'):
+        current_frame = get_video_frame(filename, frame_num)
+    else:
+        current_frame = get_image_frame(filename)
+    if current_frame is None:
+        return None, mask_top
+    
+    time.sleep(0.2)
+
+    if not fake_preview or len(roop.globals.INPUT_FACES) < 1:
+        return convert_to_gradio(current_frame), mask_top
+
+    roop.globals.face_swap_mode = translate_swap_mode(detection)
+    roop.globals.selected_enhancer = enhancer
+    roop.globals.distance_threshold = face_distance
+    roop.globals.blend_ratio = blend_ratio
+
+    if use_clip and clip_text is None or len(clip_text) < 1:
+        use_clip = False
+
+    roop.globals.execution_threads = roop.globals.CFG.max_threads
+    current_frame = live_swap(current_frame, roop.globals.face_swap_mode, use_clip, clip_text, SELECTED_INPUT_FACE_INDEX)
+    if current_frame is None:
+        return None, mask_top 
+    return convert_to_gradio(current_frame), mask_top
+
+
+def gen_processing_text(start, end):
+    return f'Processing frame range [{start} - {end}]'
+
+def on_set_frame(sender:str, frame_num):
+    global selected_preview_index, list_files_process
+    
+    idx = selected_preview_index
+    if list_files_process[idx].endframe == 0:
+        return gen_processing_text(0,0)
+    
+    start = list_files_process[idx].startframe
+    end = list_files_process[idx].endframe
+    if sender.lower().endswith('start'):
+        list_files_process[idx].startframe = min(frame_num, end)
+    else:
+        list_files_process[idx].endframe = max(frame_num, start)
+    
+    return gen_processing_text(list_files_process[idx].startframe,list_files_process[idx].endframe)
+    
+
+
+def on_preview_mask(frame_num, files, clip_text):
+    from roop.core import preview_mask
+    global is_processing
+
+    if is_processing:
+        return None
+        
+    filename = files[selected_preview_index].name
+    if util.is_video(filename) or filename.lower().endswith('gif'):
+        current_frame = get_video_frame(filename, frame_num)
+    else:
+        current_frame = get_image_frame(filename)
+    if current_frame is None:
+        return None
+
+    current_frame = preview_mask(current_frame, clip_text)
+    return convert_to_gradio(current_frame)
+
+
+def on_clear_input_faces():
+    global input_thumbs
+    
+    input_thumbs.clear()
+    roop.globals.INPUT_FACES.clear()
+    return input_thumbs
+
+def on_clear_destfiles():
+    global target_thumbs
+
+    roop.globals.TARGET_FACES.clear()
+    target_thumbs.clear()
+    return target_thumbs    
+
+
+
+def translate_swap_mode(dropdown_text):
+    if dropdown_text == "Selected face":
+        return "selected"
+    elif dropdown_text == "First found":
+        return "first"
+    elif dropdown_text == "All female":
+        return "all_female"
+    elif dropdown_text == "All male":
+        return "all_male"
+    
+    return "all"
+        
+
+
+def start_swap( enhancer, detection, keep_frames, skip_audio, face_distance, blend_ratio,
+                use_clip, clip_text, processing_method, progress=gr.Progress(track_tqdm=False)):
+    from roop.core import batch_process
+    global is_processing, list_files_process
+
+    if list_files_process is None or len(list_files_process) <= 0:
+        return gr.Button.update(variant="primary"), None
+    
+    if roop.globals.CFG.clear_output:
+        shutil.rmtree(roop.globals.output_path)
+
+
+    prepare_environment()
+
+    roop.globals.selected_enhancer = enhancer
+    roop.globals.target_path = None
+    roop.globals.distance_threshold = face_distance
+    roop.globals.blend_ratio = blend_ratio
+    roop.globals.keep_frames = keep_frames
+    roop.globals.skip_audio = skip_audio
+    roop.globals.face_swap_mode = translate_swap_mode(detection)
+    if use_clip and clip_text is None or len(clip_text) < 1:
+        use_clip = False
+    
+    if roop.globals.face_swap_mode == 'selected':
+        if len(roop.globals.TARGET_FACES) < 1:
+            gr.Error('No Target Face selected!')
+            return gr.Button.update(variant="primary"), None
+
+    is_processing = True            
+    yield gr.Button.update(variant="secondary"), None
+    roop.globals.execution_threads = roop.globals.CFG.max_threads
+    roop.globals.video_encoder = roop.globals.CFG.output_video_codec
+    roop.globals.video_quality = roop.globals.CFG.video_quality
+    roop.globals.max_memory = roop.globals.CFG.memory_limit if roop.globals.CFG.memory_limit > 0 else None
+
+    batch_process(list_files_process, use_clip, clip_text, processing_method == "In-Memory", progress)
+    is_processing = False
+    outdir = pathlib.Path(roop.globals.output_path)
+    outfiles = [item for item in outdir.rglob("*") if item.is_file()]
+    if len(outfiles) > 0:
+        yield gr.Button.update(variant="primary"),gr.Files.update(value=outfiles)
+    else:
+        yield gr.Button.update(variant="primary"),None
+
+
+def stop_swap():
+    roop.globals.processing = False
+    gr.Info('Aborting processing - please wait for the remaining threads to be stopped')
+
+
+def on_fps_changed(fps):
+    global selected_preview_index, list_files_process
+
+    if len(list_files_process) < 1 or list_files_process[selected_preview_index].endframe < 1:
+        return
+    list_files_process[selected_preview_index].fps = fps
+
+
+def on_destfiles_changed(destfiles):
+    global selected_preview_index, list_files_process
+
+    if destfiles is None or len(destfiles) < 1:
+        list_files_process.clear()
+        return gr.Slider.update(value=0, maximum=0), ''
+    
+    for f in destfiles:
+        list_files_process.append(ProcessEntry(f.name, 0,0, 0))
+
+    selected_preview_index = 0
+    idx = selected_preview_index    
+    
+    filename = list_files_process[idx].filename
+    
+    if util.is_video(filename) or filename.lower().endswith('gif'):
+        total_frames = get_video_frame_total(filename)
+    else:
+        total_frames = 0
+    list_files_process[idx].endframe = total_frames
+    if total_frames > 0:
+        return gr.Slider.update(value=0, maximum=total_frames), gen_processing_text(list_files_process[idx].startframe,list_files_process[idx].endframe)
+    return gr.Slider.update(value=0, maximum=total_frames), ''
+    
+
+
+
+def on_destfiles_selected(evt: gr.SelectData):
+    global selected_preview_index, list_files_process
+
+    if evt is not None:
+        selected_preview_index = evt.index
+    idx = selected_preview_index    
+    filename = list_files_process[idx].filename
+    fps = list_files_process[idx].fps
+    if util.is_video(filename) or filename.lower().endswith('gif'):
+        total_frames = get_video_frame_total(filename)
+        if list_files_process[idx].endframe == 0:
+            list_files_process[idx].endframe = total_frames 
+    else:
+        total_frames = 0
+    
+    if total_frames > 0:
+        return gr.Slider.update(value=list_files_process[idx].startframe, maximum=total_frames), gen_processing_text(list_files_process[idx].startframe,list_files_process[idx].endframe), fps
+    return gr.Slider.update(value=0, maximum=total_frames), gen_processing_text(0,0), fps
+    
+    
+    
+
+def on_resultfiles_selected(evt: gr.SelectData, files):
+    selected_index = evt.index
+    filename = files[selected_index].name
+    if util.is_video(filename) or filename.lower().endswith('gif'):
+        current_frame = get_video_frame(filename, 0)
+    else:
+        current_frame = get_image_frame(filename)
+    return convert_to_gradio(current_frame)
+
+
+def on_resultfiles_finished(files):
+    selected_index = 0
+    if files is None or len(files) < 1:
+        return None
+    
+    filename = files[selected_index].name
+    if util.is_video(filename) or filename.lower().endswith('gif'):
+        current_frame = get_video_frame(filename, 0)
+    else:
+        current_frame = get_image_frame(filename)
+    return convert_to_gradio(current_frame)
+
+
+    
+        
+def on_cam_toggle(state):
+    from threading import Thread
+    from roop.virtualcam import virtualcamera, cam_active
+    global live_cam_active, restart_server, camthread
+
+    live_cam_active = state
+    gr.Warning('Server will be restarted for this change!')
+    restart_server = True
+
+def on_vcam_toggle(state, num):
+    from roop.virtualcam import stop_virtual_cam, start_virtual_cam
+
+    if state:
+        start_virtual_cam(num)
+        return gr.Webcam.update(interactive=False), None
+    else:
+        stop_virtual_cam()
+    return gr.Webcam.update(interactive=True, mirror_webcam=True), None
+
+
+
+def on_stream_swap_cam(camimage, enhancer, blend_ratio):
+    from roop.core import live_swap
+    global current_cam_image, cam_counter, cam_swapping, fake_cam_image, SELECTED_INPUT_FACE_INDEX
+
+    roop.globals.selected_enhancer = enhancer
+    roop.globals.blend_ratio = blend_ratio
+
+    if not cam_swapping:
+        cam_swapping = True
+        if len(roop.globals.INPUT_FACES) > 0:
+            current_cam_image = live_swap(camimage, "all", False, None, SELECTED_INPUT_FACE_INDEX)
+        else:
+            current_cam_image = camimage
+        cam_swapping = False
+    return current_cam_image
+
+
+def on_cut_video(files, cut_start_frame, cut_end_frame):
+    if files is None:
+        return None
+    
+    resultfiles = []
+    for tf in files:
+        f = tf.name
+        destfile = util.get_destfilename_from_path(f, roop.globals.output_path, '_cut')
+        util.cut_video(f, destfile, cut_start_frame, cut_end_frame)
+        if os.path.isfile(destfile):
+            resultfiles.append(destfile)
+        else:
+            gr.Error('Cutting video failed!')
+    return resultfiles
+
+def on_join_videos(files):
+    if files is None:
+        return None
+    
+    filenames = []
+    for f in files:
+        filenames.append(f.name)
+    destfile = util.get_destfilename_from_path(filenames[0], roop.globals.output_path, '_join')        
+    util.join_videos(filenames, destfile)
+    resultfiles = []
+    if os.path.isfile(destfile):
+        resultfiles.append(destfile)
+    else:
+        gr.Error('Joining videos failed!')
+    return resultfiles
+
+
+
+
+def on_extract_frames(files):
+    if files is None:
+        return None
+    
+    resultfiles = []
+    for tf in files:
+        f = tf.name
+        resfolder = util.extract_frames(f)
+        for file in os.listdir(resfolder):
+            outfile = os.path.join(resfolder, file)
+            if os.path.isfile(outfile):
+                resultfiles.append(outfile)
+    return resultfiles
+
+
+def on_create_gif(files):
+    if files is None:
+        return None
+    
+    for tf in files:
+        f = tf.name
+        gifname = util.get_destfilename_from_path(f, './output', '.gif')
+        util.create_gif_from_video(f, gifname)
+    return gifname
+
+
+
+
+
+def clean_temp():
+    global input_thumbs, target_thumbs
+    
+    shutil.rmtree(os.environ["TEMP"])
+    prepare_environment()
+   
+    input_thumbs.clear()
+    roop.globals.INPUT_FACES.clear()
+    roop.globals.TARGET_FACES.clear()
+    target_thumbs = []
+    gr.Info('Temp Files removed')
+    return None,None,None,None
+
+
+def apply_settings(themes, input_server_name, input_server_port, output_template):
+    roop.globals.CFG.selected_theme = themes
+    roop.globals.CFG.server_name = input_server_name
+    roop.globals.CFG.server_port = input_server_port
+    roop.globals.CFG.output_template = output_template
+    roop.globals.CFG.save()
+    show_msg('Settings saved')
+
+
+def restart():
+    global restart_server
+    restart_server = True
+
+
+def show_msg(msg: str):
+    gr.Info(msg)
+
+
+
+# Gradio wants Images in RGB
+def convert_to_gradio(image):
+    if image is None:
+        return None
+    return cv2.cvtColor(image, cv2.COLOR_BGR2RGB)

roop/utilities.py

@@ -0,0 +1,343 @@
+import glob
+import mimetypes
+import os
+import platform
+import shutil
+import ssl
+import subprocess
+import sys
+import urllib
+import torch
+import gradio
+import tempfile
+import cv2
+
+from pathlib import Path
+from typing import List, Any
+from tqdm import tqdm
+from scipy.spatial import distance
+
+import roop.template_parser as template_parser
+
+import roop.globals
+
+TEMP_FILE = 'temp.mp4'
+TEMP_DIRECTORY = 'temp'
+
+# monkey patch ssl for mac
+if platform.system().lower() == 'darwin':
+    ssl._create_default_https_context = ssl._create_unverified_context
+
+
+def run_ffmpeg(args: List[str]) -> bool:
+    commands = ['ffmpeg', '-hide_banner', '-hwaccel', 'auto', '-y', '-loglevel', roop.globals.log_level]
+    commands.extend(args)
+    print (" ".join(commands))
+    try:
+        subprocess.check_output(commands, stderr=subprocess.STDOUT)
+        return True
+    except Exception:
+        pass
+    return False
+
+
+# https://github.com/facefusion/facefusion/blob/master/facefusion
+def detect_fps(target_path : str) -> float:
+    fps = 24.0
+    cap = cv2.VideoCapture(target_path)
+    if cap.isOpened():
+        fps = cap.get(cv2.CAP_PROP_FPS)
+    cap.release()
+    return fps
+
+
+	# commands = [ 'ffprobe', '-v', 'error', '-select_streams', 'v:0', '-show_entries', 'stream=r_frame_rate', '-of', 'json', target_path ]
+	# output = subprocess.check_output(commands).decode().strip()
+	# try:
+	# 	entries = json.loads(output)
+	# 	for stream in entries.get('streams'):
+	# 		numerator, denominator = map(int, stream.get('r_frame_rate').split('/'))
+	# 		return numerator / denominator
+	# 	return None
+	# except (ValueError, ZeroDivisionError):
+	# 	return 24
+
+
+def cut_video(original_video: str, cut_video: str, start_frame: int, end_frame: int):
+    fps = detect_fps(original_video)
+    start_time = start_frame / fps
+    num_frames = end_frame - start_frame
+
+    run_ffmpeg(['-ss',  str(start_time), '-i', original_video, '-c:v', roop.globals.video_encoder, '-c:a', 'aac', '-frames:v', str(num_frames), cut_video])
+
+def join_videos(videos: List[str], dest_filename: str):
+    inputs = []
+    filter = ''
+    for i,v in enumerate(videos):
+        inputs.append('-i')
+        inputs.append(v)
+        filter += f'[{i}:v:0][{i}:a:0]'
+    run_ffmpeg([" ".join(inputs), '-filter_complex', f'"{filter}concat=n={len(videos)}:v=1:a=1[outv][outa]"', '-map', '"[outv]"', '-map', '"[outa]"', dest_filename])    
+
+# def extract_frames(target_path: str) -> None:
+#     create_temp(target_path)
+#     temp_directory_path = get_temp_directory_path(target_path)
+#     run_ffmpeg(['-i', target_path, '-q:v', '1', '-pix_fmt', 'rgb24', os.path.join(temp_directory_path, f'%04d.{roop.globals.CFG.output_image_format}')])
+#     return temp_directory_path
+
+
+def extract_frames(target_path : str, trim_frame_start, trim_frame_end, fps : float) -> bool:
+    create_temp(target_path)
+    temp_directory_path = get_temp_directory_path(target_path)
+    commands = ['-i', target_path, '-q:v', '1', '-pix_fmt', 'rgb24', ]
+    if trim_frame_start is not None and trim_frame_end is not None:
+        commands.extend([ '-vf', 'trim=start_frame=' + str(trim_frame_start) + ':end_frame=' + str(trim_frame_end) + ',fps=' + str(fps) ])
+    # elif trim_frame_start is not None:
+	#     commands.extend([ '-vf', 'trim=start_frame=' + str(trim_frame_start) + ',fps=' + str(fps) ])
+    # elif trim_frame_end is not None:
+    # #     commands.extend([ '-vf', 'trim=end_frame=' + str(trim_frame_end) + ',fps=' + str(fps) ])
+    # else:
+	# 	commands.extend([ '-vf', 'fps=' + str(fps) ])
+    commands.extend([os.path.join(temp_directory_path, '%04d.' + roop.globals.CFG.output_image_format)])
+    return run_ffmpeg(commands)
+
+def create_video(target_path: str, dest_filename: str, fps: float = 24.0) -> None:
+    temp_directory_path = get_temp_directory_path(target_path)
+    run_ffmpeg(['-r', str(fps), '-i', os.path.join(temp_directory_path, f'%04d.{roop.globals.CFG.output_image_format}'), '-c:v', roop.globals.video_encoder, '-crf', str(roop.globals.video_quality), '-pix_fmt', 'yuv420p', '-vf', 'colorspace=bt709:iall=bt601-6-625:fast=1', '-y', dest_filename])
+    return dest_filename
+
+
+def create_gif_from_video(video_path: str, gif_path):
+    from roop.capturer import get_video_frame
+
+    fps = detect_fps(video_path)
+    frame = get_video_frame(video_path)
+
+    run_ffmpeg(['-i', video_path, '-vf', f'fps={fps},scale={frame.shape[0]}:-1:flags=lanczos,split[s0][s1];[s0]palettegen[p];[s1][p]paletteuse', '-loop', '0', gif_path])
+
+
+def restore_audio(intermediate_video: str, original_video: str, trim_frame_start, trim_frame_end, final_video : str) -> None:
+	fps = detect_fps(original_video)
+	commands = [ '-i', intermediate_video, '-i', original_video ]
+	if trim_frame_start is None and trim_frame_end is None:
+		commands.extend([ '-c:a', 'copy' ])
+	else:
+		if trim_frame_start is not None:
+			start_time = trim_frame_start / fps
+			commands.extend([ '-ss', format(start_time, ".2f")])
+		else:
+			commands.extend([ '-ss', '0' ])
+		if trim_frame_end is not None:
+			end_time = trim_frame_end / fps
+			commands.extend([ '-to', format(end_time, ".2f")])
+		commands.extend([ '-c:a', 'aac' ])
+	commands.extend([ '-map', '0:v:0', '-map', '1:a:0', '-y', final_video ])
+	run_ffmpeg(commands)
+
+
+
+def get_temp_frame_paths(target_path: str) -> List[str]:
+    temp_directory_path = get_temp_directory_path(target_path)
+    return glob.glob((os.path.join(glob.escape(temp_directory_path), f'*.{roop.globals.CFG.output_image_format}')))
+
+
+def get_temp_directory_path(target_path: str) -> str:
+    target_name, _ = os.path.splitext(os.path.basename(target_path))
+    target_directory_path = os.path.dirname(target_path)
+    return os.path.join(target_directory_path, TEMP_DIRECTORY, target_name)
+
+
+def get_temp_output_path(target_path: str) -> str:
+    temp_directory_path = get_temp_directory_path(target_path)
+    return os.path.join(temp_directory_path, TEMP_FILE)
+
+
+def normalize_output_path(source_path: str, target_path: str, output_path: str) -> Any:
+    if source_path and target_path:
+        source_name, _ = os.path.splitext(os.path.basename(source_path))
+        target_name, target_extension = os.path.splitext(os.path.basename(target_path))
+        if os.path.isdir(output_path):
+            return os.path.join(output_path, source_name + '-' + target_name + target_extension)
+    return output_path
+
+
+def get_destfilename_from_path(srcfilepath: str, destfilepath: str, extension: str) -> str:
+    fn, ext = os.path.splitext(os.path.basename(srcfilepath))
+    if '.' in extension:
+        return os.path.join(destfilepath, f'{fn}{extension}')
+    return os.path.join(destfilepath, f'{fn}{extension}{ext}')
+
+def replace_template(file_path: str, index: int = 0):
+    fn, ext = os.path.splitext(os.path.basename(file_path))
+    
+    # Remove the "__temp" placeholder that was used as a temporary filename
+    fn = fn.replace('__temp', '')
+
+    template = roop.globals.CFG.output_template
+    replaced_filename = template_parser.parse(template, {
+        'index': str(index),
+        'file': fn
+    })
+
+    return os.path.join(roop.globals.output_path, f'{replaced_filename}{ext}')
+
+
+
+def create_temp(target_path: str) -> None:
+    temp_directory_path = get_temp_directory_path(target_path)
+    Path(temp_directory_path).mkdir(parents=True, exist_ok=True)
+
+
+def move_temp(target_path: str, output_path: str) -> None:
+    temp_output_path = get_temp_output_path(target_path)
+    if os.path.isfile(temp_output_path):
+        if os.path.isfile(output_path):
+            os.remove(output_path)
+        shutil.move(temp_output_path, output_path)
+
+
+def clean_temp(target_path: str) -> None:
+    temp_directory_path = get_temp_directory_path(target_path)
+    parent_directory_path = os.path.dirname(temp_directory_path)
+    if not roop.globals.keep_frames and os.path.isdir(temp_directory_path):
+        shutil.rmtree(temp_directory_path)
+    if os.path.exists(parent_directory_path) and not os.listdir(parent_directory_path):
+        os.rmdir(parent_directory_path)
+
+def delete_temp_frames(filename: str) -> None:
+    dir = os.path.dirname(os.path.dirname(filename))
+    shutil.rmtree(dir)
+ 
+
+
+
+def has_image_extension(image_path: str) -> bool:
+    return image_path.lower().endswith(('png', 'jpg', 'jpeg', 'webp'))
+
+def has_extension(filepath: str, extensions: List[str]) -> bool:
+    return filepath.lower().endswith(tuple(extensions))
+
+
+def is_image(image_path: str) -> bool:
+    if image_path and os.path.isfile(image_path):
+        mimetype, _ = mimetypes.guess_type(image_path)
+        return bool(mimetype and mimetype.startswith('image/'))
+    return False
+
+
+def is_video(video_path: str) -> bool:
+    if video_path and os.path.isfile(video_path):
+        mimetype, _ = mimetypes.guess_type(video_path)
+        return bool(mimetype and mimetype.startswith('video/'))
+    return False
+
+
+def conditional_download(download_directory_path: str, urls: List[str]) -> None:
+    if not os.path.exists(download_directory_path):
+        os.makedirs(download_directory_path)
+    for url in urls:
+        download_file_path = os.path.join(download_directory_path, os.path.basename(url))
+        if not os.path.exists(download_file_path):
+            request = urllib.request.urlopen(url) # type: ignore[attr-defined]
+            total = int(request.headers.get('Content-Length', 0))
+            with tqdm(total=total, desc=f'Downloading {url}', unit='B', unit_scale=True, unit_divisor=1024) as progress:
+                urllib.request.urlretrieve(url, download_file_path, reporthook=lambda count, block_size, total_size: progress.update(block_size)) # type: ignore[attr-defined]
+
+
+def resolve_relative_path(path: str) -> str:
+    return os.path.abspath(os.path.join(os.path.dirname(__file__), path))
+
+def get_device() -> str:
+    if len(roop.globals.execution_providers) < 1:
+        roop.globals.execution_providers = ['CPUExecutionProvider']
+
+    prov = roop.globals.execution_providers[0]
+    if 'CUDAExecutionProvider' == prov:
+        return 'cuda'
+    if 'CoreMLExecutionProvider' == prov:
+        return 'mps'
+    return 'cpu'
+    
+
+def str_to_class(module_name, class_name):
+    from importlib import import_module
+    try:
+        module_ = import_module(module_name)
+        try:
+            class_ = getattr(module_, class_name)()
+        except AttributeError:
+            print('Class does not exist')
+    except ImportError:
+        print('Module does not exist')
+    return class_ or None
+
+
+# Taken from https://stackoverflow.com/a/68842705
+def get_platform():
+    if sys.platform == 'linux':
+        try:
+            proc_version = open('/proc/version').read()
+            if 'Microsoft' in proc_version:
+                return 'wsl'
+        except:
+            pass
+    return sys.platform
+
+def open_with_default_app(filename):
+    if filename == None:
+        return
+    platform = get_platform()
+    if platform == 'darwin':
+        subprocess.call(('open', filename))
+    elif platform in ['win64', 'win32']:
+        os.startfile(filename.replace('/','\\'))
+    elif platform == 'wsl':
+        subprocess.call('cmd.exe /C start'.split() + [filename])
+    else:                                   # linux variants
+        subprocess.call('xdg-open', filename)
+
+def prepare_for_batch(target_files):
+    print("Preparing temp files")
+    tempfolder = os.path.join(tempfile.gettempdir(), "rooptmp")
+    if os.path.exists(tempfolder):
+        shutil.rmtree(tempfolder)
+    Path(tempfolder).mkdir(parents=True, exist_ok=True)
+    for f in target_files:
+        newname = os.path.basename(f.name)
+        shutil.move(f.name, os.path.join(tempfolder, newname))
+    return tempfolder
+
+
+def open_folder(path:str):
+    platform = get_platform()
+    try:
+        if platform == 'darwin':
+            subprocess.call(('open', path))
+        elif platform in ['win64', 'win32']:
+            open_with_default_app(path)
+        elif platform == 'wsl':
+            subprocess.call('cmd.exe /C start'.split() + [path])
+        else:                                   # linux variants
+            subprocess.call('xdg-open', path)
+    except Exception as e:
+        print(e)
+        pass
+        #import webbrowser
+        #webbrowser.open(url)
+
+    
+
+def create_version_html():
+    python_version = ".".join([str(x) for x in sys.version_info[0:3]])
+    versions_html = f"""
+python: <span title="{sys.version}">{python_version}</span>
+•
+torch: {getattr(torch, '__long_version__',torch.__version__)}
+•
+gradio: {gradio.__version__}
+"""
+    return versions_html
+
+
+def compute_cosine_distance(emb1, emb2):
+    return distance.cosine(emb1, emb2)

roop/virtualcam.py

@@ -0,0 +1,73 @@
+import cv2
+import roop.globals
+import pyvirtualcam
+import threading
+import time
+
+
+cam_active = False
+cam_thread = None
+vcam = None
+
+def virtualcamera(cam_num):
+    from roop.core import live_swap
+
+    time.sleep(2)
+    print('Starting capture')
+    cap = cv2.VideoCapture(cam_num, cv2.CAP_DSHOW)
+    if not cap.isOpened():
+        print("Cannot open camera")
+        cap.release()
+        del cap
+        return
+
+    pref_width = 1280
+    pref_height = 720
+    pref_fps_in = 10
+    cap.set(cv2.CAP_PROP_FRAME_WIDTH, pref_width)
+    cap.set(cv2.CAP_PROP_FRAME_HEIGHT, pref_height)
+    cap.set(cv2.CAP_PROP_FPS, pref_fps_in)
+    print('Starting VCAM')
+    cam_active = True
+
+    # native format UYVY
+
+    with pyvirtualcam.Camera(width=pref_width, height=pref_height, fps=10, fmt=pyvirtualcam.PixelFormat.BGR, print_fps=True) as cam:
+
+        print(f'Using virtual camera: {cam.device}')
+        print(f'Using {cam.native_fmt}')
+
+         # RGB
+
+        while cam_active:
+            ret, frame = cap.read()
+            if not ret:
+                break
+
+            if len(roop.globals.INPUT_FACES) > 0:
+                frame = live_swap(frame, "all", False, None)
+                cam.send(frame)
+            else:
+                cam.send(frame)
+                cam.sleep_until_next_frame()
+
+    cap.release()
+    print('End cam')
+
+
+
+def start_virtual_cam(cam_number):
+    global cam_thread, cam_active
+
+    if not cam_active:
+        cam_thread = threading.Thread(target=virtualcamera, args=[cam_number])
+        cam_thread.start()
+
+
+
+def stop_virtual_cam():
+    global cam_active
+
+    cam_active = False
+    
+

run.py

@@ -0,0 +1,6 @@
+#!/usr/bin/env python3
+
+from roop import core
+
+if __name__ == '__main__':
+    core.run()

settings.py

@@ -0,0 +1,66 @@
+import yaml
+
+class Settings:
+    def __init__(self, config_file):
+        self.config_file = config_file
+        self.load()
+
+    def default_get(_, data, name, default):
+        value = default
+        try:
+            value = data.get(name, default)
+        except:
+            pass
+        return value
+
+
+    def load(self):
+        try:
+            with open(self.config_file, 'r') as f:
+                data = yaml.load(f, Loader=yaml.FullLoader)
+        except:
+            data = None
+
+        self.selected_theme = self.default_get(data, 'selected_theme', "Default")
+        self.server_name = self.default_get(data, 'server_name', "")
+        self.server_port = self.default_get(data, 'server_port', 0)
+        self.server_share = self.default_get(data, 'server_share', False)
+        self.output_image_format = self.default_get(data, 'output_image_format', 'png')
+        self.output_video_format = self.default_get(data, 'output_video_format', 'mp4')
+        self.output_video_codec = self.default_get(data, 'output_video_codec', 'libx264')
+        self.video_quality = self.default_get(data, 'video_quality', 14)
+        self.clear_output = self.default_get(data, 'clear_output', True)
+        self.live_cam_start_active = self.default_get(data, 'live_cam_start_active', False)
+        self.max_threads = self.default_get(data, 'max_threads', 2)
+        self.memory_limit = self.default_get(data, 'memory_limit', 0)
+        self.provider = self.default_get(data, 'provider', 'cuda')
+        self.force_cpu = self.default_get(data, 'force_cpu', False)
+        self.output_template = self.default_get(data, 'output_template', '{file}_{time}')
+
+
+
+
+    def save(self):
+        data = {
+            'selected_theme': self.selected_theme,
+            'server_name': self.server_name,
+            'server_port': self.server_port,
+            'server_share': self.server_share,
+            'output_image_format' : self.output_image_format,
+            'output_video_format' : self.output_video_format,
+            'output_video_codec' : self.output_video_codec,
+            'video_quality' : self.video_quality,
+            'clear_output' : self.clear_output,
+            'live_cam_start_active' : self.live_cam_start_active,
+            'max_threads' : self.max_threads,
+            'memory_limit' : self.memory_limit,
+            'frame_buffer_size' : self.frame_buffer_size,
+            'provider' : self.provider,
+            'force_cpu' : self.force_cpu,
+			'output_template' : self.output_template
+        }
+        with open(self.config_file, 'w') as f:
+            yaml.dump(data, f)
+
+
+