text
stringlengths
0
89.3k
and many reference views so that all views collectively pro
vide contextual information As illustrated in Fig 2 the key
principles in the design of a surrounding view are 1 the
main view that we focus on in this generation should oc
cupy a large proportion and 2 it should include as many
reference views as possible at a reasonable size to provide
context In practice we construct a surrounding view wrt
a specific main view by surrounding a large image of this
view with 4k1small reference images of other views
leaving a margin of arbitrary color This ensures that the
main image is roughly k2times larger than the small
images Here kis a hyperparameter The reference images
are randomly selected from all the views or nearby views
from the main view providing both a global picture of the
scene and much overlapped content to benefit training
We use such surrounding views as input images to
2 by constructing the surrounding views of the current
NeRFs rendering results structured noise and original
views Though not directly trained with this image for
mat 2 still supports generating edited images in the same
format with each image corresponding to the edited result
of the image in the same position The attention modules
in its UNet implicitly connect the same regions in differ
ent views enabling the small views to provide extra context
to the main view This results in consistently edited styles
among all the subimages in the surrounding view image
The surrounding views not only provide a contextrich
input format for 2D diffusion models but also allow it to
generate edited results for 4k3views in one batch ben
efiting our consistencyenforcing training in Sec 3333 ConsistencyEnforcing Training
We design consistent perscene training based on structured
noise and surrounding views enforcing 2D diffusion to gen
erate 3D consistent images through a consistent denoising
procedure
MultiGPU Parallelization Paradigm Our pipeline in
volves training both NeRF and 2D diffusion Observing
that training and inferring a diffusion model is consider
ably more timeconsuming than training the NeRF while
there are very few dependencies between them we propose
a multiGPU parallelization paradigm With n 1 GPUs
we dedicate GPU 0to continuously and asynchronously
train a NeRF on the buffer of edited images The remain
ingnGPUs are utilized to train the diffusion model and
generate new edited images added to the buffer for NeRF
training At each diffusion training iteration we allocate
a view to each of the nGPUs and train diffusion on them
synchronously This parallelization eliminates the need to
explicitly trade off between NeRF and diffusion training
leading to a 10speedup in training With multiple dif
fusion generations running synchronously we can also en
force crossgeneration consistency
Augmenting 2D Diffusion with 3DInforming Control
Net Intuitively a 3Dconsistent model needs to be 3D
aware otherwise it lacks the necessary information and
may solely adapt to the input structured noise potentially
leading to overfitting Therefore we incorporate an addi
tional ControlNet 47 adaptor into our 2D diffusion which
injects 3D information as a new condition The 3D infor
mation is obtained by using NeRF to infer the depth and
3D point for each pixel in the view We then query its fea
ture in a learnable 3D embedding implemented as a hash
table in 23 to acquire a pixelwise 3Daware feature im
age which serves as the condition for ControlNet These
components make the augmented diffusion to be aware of
learn from and generate results based on 3D information
Additionally we apply LoRA 11 to further enhance the
capability of diffusion
SelfSupervised Consistency Loss Lacking ground
truth for consistently edited images we introduce a self
supervised method to enforce 3D consistency For a set of
generated multiview images we construct a corresponding
reference set of 3D consistent multiview images to serve
as a selfsupervision target Inspired by 37 42 we em
ploy depthbased warping with NeRFrendered depth to es
tablish pixel correspondence across views We design a
weighted averaging process to aggregate these pixels to the
final image ensuring multiview consistency detail in sup
plementary
Specifically we edit nsurrounding views synchronously
onnGPUs with each surrounding view containing 4k3
views resulting in a total of V 4k3nviews For each
viewv we warp the edited results of the remaining V1
5Figure 3 Comparison in the Fangzhou scene shows that our ConsistDreamer produces significantly sharper editing results with more
finegrained textures and higher consistency with the instruction eg Lord V oldemort with no hair on which all baselines fail The
instructions are the bottom texts except for NArt 34 which uses the underlined texts The images of baselines are taken from their paper
views to it and compute their weighted average to obtain
the reference view v We then reaggregate reference views
vback into surrounding views in the original structure
for each GPU These reassembled surrounding views are
then used as the target images to supervise 2D diffusion
To guide 2D diffusion in preserving the original style
and avoiding smoothing out we define our consistency loss
as the sum of the VGGbased perceptual and stylization
loss 12 instead of a pixelwise loss between diffusions
output and the target image In addition to this primary loss
we propose several regularization losses to prevent mode
collapse and promote 3D awareness detail in supplemen
tary With the consistency loss ConsistDreamer effec
tively enforces not only crossview consistency among all
views in each surrounding view but also crossgeneration or