text
stringlengths 0
89.3k
|
---|
core contributions structured noise surrounding views |
and consistencyenforcing training can also be extended to |
the scene generation task For example these components |
can be used in the refinement phase when the shape of a |
scene is roughly determined In this way these components |
could help achieve consistent and highfidelity generation |
refining the shape with slight adjustments for more detailed |
and precise geometry Compared with previous methods |
18 36 this method can generate scenes with detailed |
highfidelity textures and shapes without mesh exportation |
or fixing geometry |
G2 Limitations |
This section discusses the limitations of ConsistDreamer |
which are also the common challenges encountered by ex |
isting 3D scene editing methods |
ViewDependent or Specular Effects Our Consist |
Dreamer pipeline performs consistencyenforcing training |
by warping and averaging between different views This |
procedure enforces that each part of the scene looks the |
same in different views ie is viewindependent mak |
ing the edited scene unlikely to show viewdependent or |
specular effects To preserve the ability to generate view |
dependent effects our ConsistDreamer has introduced a |
regularization loss that trades off between consistency and |
similarity to original 2 detailed in Sec D8 With this |
regularization our ConsistDreamer could still achieve 3D |
consistency while allowing natural viewdependent effects |
The baselines though are not trained towards consistencyor viewindependence only generate blurred results with |
out notable effects or even overfit to inconsistent editing |
with the viewdependency of NeRF |
Editing Capabilities Constrained by 2D Diffusion Mod |
els Our ConsistDreamer distills from the diffusion model |
2 to edit scenes Therefore the editing ability style and |
diversity of ConsistDreamer are inherently constrained by |
2 Our ConsistDreamer edits a scene in a specific man |
ner following 2 For example in the Vincent Van Gogh |
editing in Fig B2 our ConsistDreamer along with IN2N |
8 and ViCA 6 which use the same 2 for editing shows |
a side effect that transfers the style of the image to Van |
Goghs painting style Moreover we cannot support edit |
ing tasks on which the diffusion model cannot perform |
Despite this common constraint among all the distillation |
based methods our ConsistDreamer successfully transfers |
most of the editing capabilities of the 2D diffusion model to |
3D by achieving highquality and highdiversity 3D scene |
editing |
3D Understanding and Reasoning Though our Consist |
Dreamer is 3Dinformed and 3Daware with the additional |
input of 3D positional embedding already surpassing all |
the baselines it is unable to reason and understand the se |
mantics of each part of 3D scenes Therefore while our |
ConsistDreamer can edit a view using the knowledge of the |
whole scenes shape via 3D positional embedding and ap |
pearance through the surrounding view it may still en |
counter multiface issues or Janus problems Specifically it |
does not understand what the correct orientation of the face |
is does not know that a person can only have one face and |
thus cannot avoid this problem |
Shape Editing Some instructions for editing tasks may |
involve modifying the geometry or shape of a scene eg |
give him a beard creates a beard on the face Like the |
baselines our ConsistDreamer is designed to support sim |
ple shape editing tasks that can be achieved by slightly and |
gradually alternating the surface For example in the edit |
ing of give him a beard our pipeline gradually grows |
the beards shape from the faces surface Notice that |
both ConsistDreamer and the baselines cannot perform ag |
gressive and complicated editing eg removing an object |
while reconstructing the whole occluded part or direction |
related editing eg performing lower down her arm for |
a scene of a person raising her arm requires a multiview |
consensus on which direction the arm is moved to |
Efficiency In contrast to the diffusion trainingfree base |
lines such as 6 8 13 our ConsistDreamer needs addi |
tional training of a 2D diffusion model This extends the |
editing duration resulting in taking 12 hours to edit a scene |
in the IN2N dataset and up to 24 hours to edit a large |
scale indoor scene in the ScanNet dataset However as |
a tradeoff against efficiency our ConsistDreamer excels in |
achieving highfidelity editing surpassing all the training |
17free baselines |
G3 Future Directions |
Supporting Specular Effects One direction is to support |
specular effects and better viewdependency This may need |
an improved formulation of consistency under specular re |
flections or modeling the ambient environment |
3D Understanding for Scene Editing Another direction |
is to enable the diffusion model to understand and reason |
the semantics of a scene Introducing a model that gener |
ates 3D semantic embeddings for each point in the scene al |
lows for combining this information with the 3D positional |
embedding as the input to the diffusion model potentially |
mitigating Janus problems |
18International Journal of Computer Information Systems and Industrial Management Applications |
ISSN 21507988 V olume 142022 pp 2847 |
MIR Labs wwwmirlabsnetijcisimindexhtml |
Submitted 28 October 2021 Accepted 2 February 2022 Published 9 April 2022 |
Enhanced Elephant Herding Optimization for Large |
Scale Information Access on Social Media |
Yassine Drias1 Habiba Drias2and Ilyes Khennak3 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.