text
stringlengths
0
89.3k
applied upon a diffusionbased 2D image editing model 2
As illustrated in Fig 2 our pipeline maintains a buffer of
edited views for the NeRF to fit and uses 2 to generate
new edited images for random views according to the in
struction the original appearance and the current NeRF
rendering results Noticing that the NeRF fitting proce
dure and diffusion generation procedure are relatively inde
pendent we equivalently execute them in parallel Within
3Figure 2 ConsistDreamer framework is an IN2Nlike 8 pipeline containing two major procedures aIn the NeRF fitting procedure
we continuously train NeRF with a buffer of edited views bIn diffusion generation and training we add our 3Dconsistent structured
noise to rendered multiview images and compose surrounding views with them as input to the augmented 3Daware 2D diffusion 2
We then add the edited images to the buffer for a and apply selfsupervised consistencyenforcing training using the consistencywarped
images Note images of structured noise are only for illustration they are actually visually indistinguishable from Gaussian noise images
this framework we propose 1 structured noise to enable
a 3Dconsistent denoising step starting from 3Dconsistent
noise at the beginning and ending with 3Dconsistent im
ages 2 surrounding views to construct contextrich com
posed images as input to the 2D diffusion instead of a single
view and 3 a selfsupervised consistencyenforcing train
ingmethod via consistent warping in surrounding views to
achieve crossview and crossbatch consistency
31 Structured Noise
2D diffusion models generate a new image from a noisy
image which is either pure Gaussian noise or a mixture
of noise and the original image Prior works like Dream
Fusion 25 and IN2N typically sample different Gaussian
noise in each iteration However varying noise leads to
highly different generation results as shown in the supple
mentary material In other words previous methods cannot
even produce consistent ie identical images for the same
view in different generations fundamentally limiting their
ability to generate consistent results
This observation motivates us to control and manipulate
the noise by introducing 3Dconsistent structured noise
Intuitively while it is difficult to generate denoise or re
store 3Dconsistent images from inconsistent random noise
the task becomes more manageable when generating consis
tent images from noise that is itself consistent Therefore
instead of using independently generated noise in each it
eration we generate noise on the surface of the scene only
once during initialization and render the noise at each view
to obtain the noise used in generating the image for that
view Our strategy aligns with 3D diffusion models like
DiffRF 22 which directly generate noise in 3D space The
difference lies in the denoising step while such work di
rectly denoises in 3D we distill the 3D denoising processfrom pretrained 2D diffusion models
As a latent diffusion model 2 actually requires noise in
latent space which is H8 W84instead of the image
shape H W 3 Each element in this noise latent should be
independently generated from N01 Constructing such
3Dconsistent structured noise remains nontrivial we need
to place noise in 3D project noise into 2D pixels at multiple
scales and ensure correspondence between different views
Additionally the distribution of each images noise should
be Gaussian as noise in an incorrect or dependent distribu
tion may lead to abnormal generation results as shown in
the supplementary material
To overcome these challenges we construct a dense
point cloud of the scene by unprojecting all the pixels in
all the views to points with the depth predicted by NeRF
For each point p we randomly sample a weighted noise
cp x w where xN01is an independently
generated Gaussian noise and wU01is its weight
To generate the noise at one view we identify the sub
point cloud that is frontmost in this view and project it
onto the plane For multiple points projected to the same
pixel we aggregate them by selecting the weighted noise
x wwith the maximum w and form a noise image Iof
shape H W consisting of values in x As each xis inde
pendently generated and selected according to w we have
IN01HWie making Ivalid 2D Gaussian noise
Given that each pixel in the latent space is only roughly
related to its corresponding 88region in the image
we can generate noise in the latent space by operating at
the downsampled resolution of H8W8 We thus
generate different weighted noise cipfor each of the
four channels of the latent space and stack the individu
ally rendered noise Iito construct a Gaussian noise image
ofH8 W84 which is then used as the noise by the
4diffusion model
The structured noise serves as the foundation for 3D
consistent generation In Sec 33 we introduce a training
method to ensure a consistent denoising procedure from the
beginning to the end so that the denoised images of differ
ent views at every denoising step is also 3D consistent
32 Surrounding Views
Using the original view as input is a standard practice when
employing 2D diffusion models This method works well in
simple 360or forwardfacing scenes used by IN2N 8 as
a single view covers most objects in the scene However in
more complicated scenes like the cluttered rooms in Scan
Net 43 a view may only contain a corner or a bare wall
in the room This hinders the diffusion model to generate
plausible results due to the limited context in a single view
Intriguingly our investigation reveals that 2 performs
well on composed images generating an image composed
of styleconsistent edited subimages with the same struc
ture as shown in the supplementary This observation in
spires us to exploit a novel input format for diffusion models
surrounding views with a composition of one main view