text
stringlengths
0
89.3k
bedding at least for novel view synthesis to recover theoriginal image encouraging the utilization of 3D infor
mation This regularization loss is also applied on the
UNet in each denoising step
Encourage Consistent Editing Style The diffusion model
has some diversity in editing However we need to con
verge to one specific style in one editing procedure oth
erwise the NeRF may use viewdependency to overfit
different styles at different views Therefore in the Pre
Annealing step Sec D3 we use the NeRFs rendering
result to supervise the diffusion model to make it con
verge to the style NeRF converges to
D9 Variant IN2N And IN2N 8
In our ablation study in the main paper we have a variant
IN2N being our full ConsistDreamer with all three major
components removed In this section we discuss how it
is equivalent to an implementation of IN2N and the major
differences between them
IN2N is a method that 1 gradually generates newly
edited images with a noise level detailed in Sec D3 sam
pled from 7098 and 2 uses the newly generated
images to fit the NeRF while the fitting NeRFs rendering
results can affect the following editing through the input
of diffusion model as a mixture with noise This matches
our preannealing substage Therefore IN2N includes
vanilla IN2N as a subprocedure Additionally IN2N has
the following improvements beyond IN2N
IN2N only samples noise levels from 7098 This
makes IN2N 1 sometimes unable to sufficiently edit the
scene due to the absence of 100 noise level editing eg
unable to achieve a Lord V oldemort editing with no hair
in Fig B2 and 2 cannot refine the editing results based
on a converged style and sometimes even deviates from a
converged style to another as the noise level is always as
high as 70 The variant IN2N starts at a full noise be
fore the preannealing substage guaranteeing sufficient
editing After the preannealing substage IN2N an
neals the noise level range to refine the results leading to
a more finegrained editing
IN2N adds the newly edited image to the dataset by re
placing a subset of pixels which may negatively affect
the LPIPSperceptual loss IN2N uses an edited view
buffer to fit NeRF containing only full edited views on
which the perceptual loss can perform well
In conclusion our variant IN2N is an equivalent and im
proved implementation of IN2N As shown in SV IN2N
generates noticeably better results than IN2N
15Figure E1 The pretrained diffusion model 2 works as expected on surrounding views by editing each subview in the instructed way
individually but in a consistent style Notably as shown in the last row the surrounding view enriches the context making the diffusion
model succeed in views that fail in singleview editing
Figure E2 Even for the same view generating from different noises does not necessarily lead to the consistent ie the same edited result
Each column represents a generation from a noise different from other columns
E Supporting Evidence for Claims
E1 Diffusion Models Perform Well with Composed
Images
As shown in Fig E1 the pretrained diffusion model 2
though not directly trained in this pattern still works as ex
pected in surrounding views It generates editing results for
each subview individually while all of them also share a
similar style across various scenes including indoor out
door and faceforwarding scenes
Notably as shown in the last row when editing a view
with little context directly editing the single view fails
Constructing a surrounding view using it as the main view
however helps the diffusion model 2 to achieve success
ful editing This shows the effects of surrounding views in
achieving successful and consistent editingE2 Different Noises Lead to Varied Results
As shown in Fig E2 generation from different noises leads
to completely different images which is the fundamental
constraint of all the baselines which do not control the
noise Even with surrounding views the diffusion model
2 still generates images in highly inconsistent ways The
diversity of the diffusion model under different noises is de
sirable in 2D generation and editing but has to be controlled
in 3D generation for consistency
F Additional Ablation Study Analysis
F1 No Str Noise vs Only Sur Views
Both variants do not have structured noise Hence the
consistencyenforcing training in No Str Noise forces
the model to generate the same result from different noises
16which leads to mode collapse and degrades the editing result
towards blurred averaged color These negative effects of
training in No Str Noise leads to similar and even worse
results and DFS than Only Sur Views with no training
F2 Only Sur Views vs IN2N
Tasks BCD are style transfer specifically well supported
by our current 2D diffusion model 2 Our DFS metric
based on FID uses a feature extractor with more tolerance
for different style transfer results in the same image Hence
even IN2N performs comparably with a slightly lower
DFS
By contrast task A is a general objectcentric editing
with diversified editing manners different valid editing re
sults can have jackets with completely different colors and
styles There can even be geometric changes in the cloth
ing without surrounding views as context to constrain the
editing leading to a significantly worse DFS for IN2N
G Discussion
G1 Extension to Scene Generation
The proposed ConsistDreamer primarily focuses on the
distillationguided 3D scene editing task However the