text
stringlengths
0
89.3k
ing all the pixels in all the views and renderingprojecting
such a point cloud at a view to generate the structured
noise Directly implementing this literal description is com
plicated and inefficient
Therefore we use an equivalent implementation
Instead of explicitly generating this dense point cloud we
just put the weighted noises on each pixel of all views
For the view we query for structured noise we warp the
noise from all other views to it This is equivalent to pro
jecting the subpoint cloud generated by each view to the
querying view therefore it is equivalent to the original
design
With this implementation explicitly generating main
taining and projecting a point cloud with billions of points
number of views height width is unnecessary and a
query can be completed in less than one second
D5 Surrounding Views Reference View Selection
We construct the surrounding view with one large main
view and several small reference views The purpose of
the reference views is twofolded 1 to provide enough
context about the whole scene and 2 to have enough
overlapped parts of the main view to facilitate consistency
enforcing training Therefore we select 40 of the views to
be a random view of the scene and the rest 60 of the views
to be a view with at least 20 overlap of the main view
quantified by the area of matched pixels through warping
The order of the views is randomly shuffled We observed
that none of these randomnesses highly alter the editing re
sult after consistencyenforcing training any choice of
reference views and their order will lead to a consistent
edited result of the main view
D6 Training Pixel Weights
In consistencyenforcing training we apply warping and
weighted averages to compute the training reference views
v so that all the views in vare 3D consistent Using
identical weights for all pixels will result in blurred images
In a scene of a person one view only contains their face
and another view contains the whole body Warping the lat
ter to the former indicates an upsampling of the face part
which will be blurred Merging the blurred warped view
with the former view at the same weight results in blurred
overall results
We propose a better pixelweighting strategy based on a
further analysis of this situation If we warp pixel ato pixel
b where ahas a larger scope and contains more scene
objects then we need to upsample the bpart of the view
froma resulting in blurry Therefore the weight should be
related to the scope of the pixel Following this we define
14the pixel area to quantify this scope For a pixel pin a view
from camera position o the four vertices of the pixel grids
correspond to the rays otdi4
i1 We use NeRF to predict
each of their depth ti4
i1 and calculate their correspond
ing points Piotidi The pixel area Spof this pixel is
defined as the area of a square with vertices P1 P2 P3 P4
in the 3D space which can be regarded as an approximation
of the surface area the pixel represents As we need a lower
weight for a pixel with a larger scope ielarger Sp we
define the weight as 1Sp which satisfies all our needs
D7 Training MultiGPU Pipeline
An illustration of our multiGPU training pipeline is in Fig
D1 By implementing such a parallelization pipeline by
ourselves we decouple NeRF training with diffusion gen
eration and training in the most asynchronized way waiving
the necessity of tradeoffs between NeRF training and dif
fusion achieving considerable speed up
D8 Training Regularizations
We use the consistency loss as the main loss in the
consistencyenforcing training However this loss only en
forces several equalities required by consistency leading
to trivial results of a purecolor image without regulariza
tion losses this is also reasonable as all purecolor images
of the same color are perfectly consistent Also there is no
encouragement or enforcement to use the 3D information in
the 3D positional embedding To avoid these we propose
several regularization losses as shown below
Maintain Original Behavior We expect that the trained
diffusion model will generate images that are very similar
to the original model when all the inputs image noises
and 3D positional embeddings are identical Therefore
we use MSE and VGG perceptual and stylization losses
to regularize both the generated images and the con
structed referenced images with gradient generated by
warping and averaging of the trained diffusion with the
original models output We further expect that the UNet
in the trained diffusion model predicts similar noises at
each denoising step as the original UNet so we also use
this to regularize during each denoising step
Encourage 3D Information Utilization The original 2
takes the original image of the scene as another part of
input using it as a condition to generate the edited image
To encourage 3D information utilization we design a reg
ularization loss to enforce the diffusion model without
the original image input to generate very similar results
to the one with the original image input both with 3D po
sitional embedding input With the lack of the original
image the only way for the diffusion model to perceive
the original view is the 3D positional embedding There
fore the diffusion is trained to use the 3D positional em