text
stringlengths
0
89.3k
the upload size limitation of 200MB on OpenReview
In the following sections we use SVto refer to this sup
plementary video
B Comparisons with Additional Baselines
In the main paper we compare our ConsistDreamer with
IN2N 8 and ViCA 6 In this section we compare
our ConsistDreamer with other baselines and provide some
analysis These methods either do not have publicly avail
able code or evaluate on the scenes which are not supported
by NeRFStudio Therefore we could only compare our
ConsistDreamer under the tasks used by them with the pro
vided visualizations from their papers or websites
We also provide some comparisons in the video format
of the baselines in SV
B1 CSD 15
CSD is a method focusing on general consistent generation
including large image editing scene editing and scene gen
eration We compare our ConsistDreamer with CSD un
der three tasks shown on the website of CSD1 LowPoly
1httpssubinkimcvgithubioCSD
Figure B2 Compared with DreamEditor our ConsistDreamer
achieves better editing which not only follows and satisfies the
given instructions but also preserves as much content of the orig
inal scene as possible On the contrary DreamEditor completely
edits the original person to another in all the tasks
Graphic Anime and Smile
As shown in Fig B1 and SV our ConsistDreamer sig
nificantly outperforms IN2N which fails in the LowPoly
and Anime tasks and has the side effects of adding beards
in the Smile task Compared with CSD our editing in
the LowPoly task is more noticeable with a successfully
edited hair part Our edited scene in the Smile task is the
only one among all three to successfully show the teeth
when smiling while CSDs result contains strange muscles
as if the person is keeping a straight face In conclusion
our ConsistDreamer achieves more successful editing than
CSD
B2 DreamEditor 49
DreamEditor is another method focusing on scene editing
but with another diffusion model 29 instead of 2 As
NeRFStudio does not support the other scenes we compare
our ConsistDreamer with DreamEditor by comparing Fig
3 in our main paper with Fig 8 in 49
11Figure B3 Our ConsistDreamer achieves consistent editing in the
checkeredplaid pattern also visualized as smooth video in SV
while EditDiffNeRF has obvious inconsistency in the shape and
texture of the collar
Fig B2 presents the results in these tasks along with
other baselines in Fig 3 in our main paper It shows that
our ConsistDreamer preserves most of the contents in the
original scene while editing eg the shape of the head and
face and the shape and type of the clothes minimizing the
side effects of editing DreamEditor however completely
edits the person to another person even in the Fauvism task
which is supposed to be only style transfer This demon
strates that our ConsistDreamer achieves more reasonable
editing than DreamEditor
B3 EditDiffNeRF 45
EditDiffNeRF is another paper that also claims to suc
cessfully complete the checkeredplaid pattern As they
did not provide any code we compare our ConsistDreamer
with the images provided in their paper As shown in Fig
B3 our ConsistDreamer achieves consistent editing among
all three views while EditDiffNeRFs results are multi
view inconsistent obviously shown in the collar part The
smooth video of our rendering result in SValso shows the
consistency of our ConsistDreamer These results validate
that our ConsistDreamer archives significantly better con
sistency in checkeredplaid patterns while EditDiffNeRF
fails to achieve such consistency
B4 Instruct 3Dto3D 13
Instruct 3Dto3D is a method focusing on style transfer of
scenes It uses LLFF and NeRF Synthetic NS scenes as
editing tasks instead of the widelyused IN2N dataset Incontrast we focus on editing more challenging and realis
tic scenes In addition as NeRFStudio and NeRFacto do
not support LLFF and NS datasets well more specifically
NeRFStudio does not support the LLFF dataset and NeR
Facto works well in real scenes but not in synthetic scenes
like NS we cannot compare with Instruct 3Dto3D on
these two datasets Moreover the code of Instruct 3Dto
3D is not publicly available Therefore we are unable to
compare with Instruct 3Dto3D
B5 Concurrent Works GE 5 EN2N 31 And
PDS 17
5 17 31 are three concurrent works GE 5 and EN2N
31 achieve 3D editing through the same 2D diffusion
model 2 and have some modifications in the pipeline or
scene representation while PDS 17 proposes another dis
tillation formula and uses DreamBooth 29 for editing
The comparisons against them are in Fig B4 Our
ConsistDreamer generates highquality editing results with
brighter color and clearer textures while all these concur
rent works generate blurred textures gloomy colors andor
unsuccessful or unreasonable editing
C CLIP 26 Metrics In IN2N 8
We provide the quantitative comparison with CLIP 26
metrics introduced in IN2N 8 in Tab C1 In all four abla
tion scenes ours significantly and consistently outperforms
IN2N in both metrics
D Implementation Details