tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
1arXiv240609404v1 csCV 13 Jun 2024most of its details and sharpness A commonly observed
failure mode is that of regular eg checkered patterns
which completely disappear once distilled to 3D due to mis
alignments across views Generating consistent multiview
images thus becomes crucial for achieving highfidelity 3D
scene editing
While largely overlooked in prior work our investiga
tion reveals that the source of inconsistency is multifaceted
and primarily originates from the input 1As the 2D dif
fusion model can only observe a single view at a time it
lacks sufficient context to understand the entire scene and
apply consistent editing 2The editing process for each
image starts from independently generated Gaussian noise
which brings challenges to consistent image generation In
tuitively it is difficult to generate consistent multiview im
ages by denoising inconsistent noise and even for a single
view it may not always yield the same edited result 3The
input to the 2D diffusion model contains no 3D information
making it much harder for the model to reason about 3D
geometry and to share information across different views of
the scene even when made available to it
Motivated by these observations we propose Consist
Dreamer a novel framework to achieve 3D consistency in
2D diffusion distillation ConsistDreamer introduces three
synergistic strategies that augment the input of the 2D dif
fusion model to be 3Daware and enforce 3D consistency in
a selfsupervised manner during the training process
To address the limited context issue within a single view
our framework involves incorporating context from other
views We capitalize on the observation that 2D diffusion
models inherently support composed images where mul
tiple subimages are tiled to form a larger image Given
the capability of the selfattention modules in the UNet of
the 2D diffusion model to establish connections between
the same objects across different subimages each image
can be edited with the context derived from other images
Therefore we leverage the composed images to construct
asurrounding view Fig 1 where one large central main
view is surrounded by several small reference views This
approach allows us to edit the main view with the context
from reference views and vice versa Doing so not only en
riches the context of the scene in the input but also enables
the simultaneous editing of multiple views
Regarding the noise we introduce 3Dconsistent struc
tured noise Fig 1 with the key insight of generating con
sistent noise for each view once at the beginning Specifi
cally we generate and fix Gaussian noise on the surface of
the scene objects and then render each view to obtain the
2D noise used for the image at that view in all subsequent
diffusion generations This approach aligns with existing
3D diffusion work 22 which also generates noise in 3D at
the beginning of a generation Ensuring that the denoising
procedure starts with consistent noise substantially facilitates the process of achieving consistent images by the end
The combination of surrounding views and structured
noise provides the 2D diffusion model with 3D consistent
input yet it is insufficient An explicit enforcement of 3D
consistency is also required during the learning process To
this end we propose selfsupervised consistencyenforcing
training within the perscene editing procedure Fig 1 We
augment the 2D diffusion model by a ControlNet 47 that
introduces 3D positional embedding to make it 3Daware
Inspired by 37 42 we perform warping and averaging for
all subviews in the edited surrounding view image This
process yields a surrounding view of 3D consistent sub
views used as the selfsupervision target To further achieve
crossbatch consistency consistency between different
batches in different generations we perform multiple gen
erations in parallel and construct consistent target images
from all subviews in all generated surrounding view im
ages so as to supervise all generations collectively Af
ter consistencyenforcing training the 2D diffusion model
is able to generate consistent multiview images Conse
quently a trained NeRF will not have to smooth out incon
sistencies but ultimately converge to sharp results preserv
ing finegrained details
Empowered by such a 3Dconsistent 2D diffusion
model our ConsistDreamer achieves highfidelity and di
verse instructionguided 3D scene editing without any mesh
exportation and refinements or a better scene representation
like Gaussian Splatting 14 as shown in Fig 1 Com
pared with previous work the editing results of Consist
Dreamer exhibit significantly improved sharpness and de
tail while preserving the diversity in the original 2D diffu
sion models 2 editing results Notably ConsistDreamer
stands as the firstwork capable of successfully editing com
plex eg checkered patterns Moreover ConsistDreamer
demonstrates superior performance in complicated high
resolution ScanNet 43 scenes an accomplishment
where stateoftheart methods faced challenges in achiev
ing satisfactory edits
Our contributions are threefold 1We introduce
ConsistDreamer a simple yet effective framework that en
ables 3Dconsistent instructionguided scene editing based
on distillation from 2D diffusion models 2We propose
three novel synergistic components structured noise sur
rounding views and consistencyenforcing training that
lift 2D diffusion models to generate 3Dconsistent images
across all generated batches Notably our work is the first
that explores crossbatch consistency and denoising consis
tency in 2D diffusion distillation and attains these through

1arXiv240609404v1 csCV 13 Jun 2024most of its details and sharpness A commonly observed

failure mode is that of regular eg checkered patterns

which completely disappear once distilled to 3D due to mis

alignments across views Generating consistent multiview

images thus becomes crucial for achieving highfidelity 3D

scene editing

While largely overlooked in prior work our investiga

tion reveals that the source of inconsistency is multifaceted

and primarily originates from the input 1As the 2D dif

fusion model can only observe a single view at a time it

lacks sufficient context to understand the entire scene and

apply consistent editing 2The editing process for each

image starts from independently generated Gaussian noise

which brings challenges to consistent image generation In

tuitively it is difficult to generate consistent multiview im

ages by denoising inconsistent noise and even for a single

view it may not always yield the same edited result 3The

input to the 2D diffusion model contains no 3D information

making it much harder for the model to reason about 3D

geometry and to share information across different views of

the scene even when made available to it

Motivated by these observations we propose Consist

Dreamer a novel framework to achieve 3D consistency in

2D diffusion distillation ConsistDreamer introduces three

synergistic strategies that augment the input of the 2D dif

fusion model to be 3Daware and enforce 3D consistency in

a selfsupervised manner during the training process

To address the limited context issue within a single view

our framework involves incorporating context from other

views We capitalize on the observation that 2D diffusion

models inherently support composed images where mul

tiple subimages are tiled to form a larger image Given

the capability of the selfattention modules in the UNet of

the 2D diffusion model to establish connections between

the same objects across different subimages each image

can be edited with the context derived from other images

Therefore we leverage the composed images to construct

asurrounding view Fig 1 where one large central main

view is surrounded by several small reference views This

approach allows us to edit the main view with the context

from reference views and vice versa Doing so not only en

riches the context of the scene in the input but also enables

the simultaneous editing of multiple views

Regarding the noise we introduce 3Dconsistent struc

tured noise Fig 1 with the key insight of generating con

sistent noise for each view once at the beginning Specifi

cally we generate and fix Gaussian noise on the surface of

the scene objects and then render each view to obtain the

2D noise used for the image at that view in all subsequent

diffusion generations This approach aligns with existing

3D diffusion work 22 which also generates noise in 3D at

the beginning of a generation Ensuring that the denoising

procedure starts with consistent noise substantially facilitates the process of achieving consistent images by the end

The combination of surrounding views and structured

noise provides the 2D diffusion model with 3D consistent

input yet it is insufficient An explicit enforcement of 3D

consistency is also required during the learning process To

this end we propose selfsupervised consistencyenforcing

training within the perscene editing procedure Fig 1 We

augment the 2D diffusion model by a ControlNet 47 that

introduces 3D positional embedding to make it 3Daware

Inspired by 37 42 we perform warping and averaging for

all subviews in the edited surrounding view image This

process yields a surrounding view of 3D consistent sub

views used as the selfsupervision target To further achieve

crossbatch consistency consistency between different

batches in different generations we perform multiple gen

erations in parallel and construct consistent target images

from all subviews in all generated surrounding view im

ages so as to supervise all generations collectively Af

ter consistencyenforcing training the 2D diffusion model

is able to generate consistent multiview images Conse

quently a trained NeRF will not have to smooth out incon

sistencies but ultimately converge to sharp results preserv

ing finegrained details

Empowered by such a 3Dconsistent 2D diffusion

model our ConsistDreamer achieves highfidelity and di

verse instructionguided 3D scene editing without any mesh

exportation and refinements or a better scene representation

like Gaussian Splatting 14 as shown in Fig 1 Com

pared with previous work the editing results of Consist

Dreamer exhibit significantly improved sharpness and de

tail while preserving the diversity in the original 2D diffu

sion models 2 editing results Notably ConsistDreamer

stands as the firstwork capable of successfully editing com

plex eg checkered patterns Moreover ConsistDreamer

demonstrates superior performance in complicated high

resolution ScanNet 43 scenes an accomplishment

where stateoftheart methods faced challenges in achiev

ing satisfactory edits

Our contributions are threefold 1We introduce

ConsistDreamer a simple yet effective framework that en

ables 3Dconsistent instructionguided scene editing based

on distillation from 2D diffusion models 2We propose

three novel synergistic components structured noise sur

rounding views and consistencyenforcing training that

lift 2D diffusion models to generate 3Dconsistent images

across all generated batches Notably our work is the first

that explores crossbatch consistency and denoising consis

tency in 2D diffusion distillation and attains these through