tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
manipulating noise 3We evaluate a range of scenes and
editing instructions achieving stateoftheart performance
in both scenes considered by previous work and more com
plicated largescale indoor scenes from ScanNet
22 Related Work
NeRFBased Scene Editing Neural radiance field
NeRF 21 and its variants 1 3 7 19 33 35 38 44
are widelyused approaches to representing scenes NeRF
leverages neural networks or other learnable architectures to
learn to reconstruct the 3D geometry of a scene only from
multiview images and their camera parameters and sup
port novel view synthesis With the development of NeRF
editing a NeRFrepresented scene is also deeply studied
covering different types of editing objectives and editing
operation indicators aka user interfaces Some meth
ods 16 20 41 support editing the position color andor
shape of a specific object indicated by users through a
pixel a text description or a segment etc Another line
of work 4 24 39 46 studies humanguided shape edit
ing which allows users to indicate a shape editing oper
ation with a cage or point cloud provided by the model
The task we investigate is instructionguided scene editing
which allows users to indicate the editing operation through
instructions in natural language The first work in this di
rection is NeRFArt 34 which mainly focuses on style
transfer instructions and uses pretrained CLIP 26 as the
stylization loss for the instructionindicated style More re
cent work 8 13 15 45 49 leverages diffusion models 10
instead of CLIP to benefit from powerful diffusion models
and support more general instructions
DistillationBased 3D Scene Generation Lacking 3D
datasets to train powerful 3D diffusion models current so
lutions distill the generation signal from a 2D diffusion
model to exploit its ability in 3D generation DreamFu
sion 25 is the first work in this direction which proposes
score distillation sampling SDS to distill the gradient up
date direction score from 2D diffusion models and
supports instructionguided scene generation by distilling a
pretrained diffusion model 28 HiFA 48 proposes an an
nealing technique and rephrases the distillation formula to
improve the generation result Magic3D 18 improves the
generation results by introducing a coarsetofine strategy
and a mesh exportation and refinement method Prolific
Dreamer 36 further improves the generation results by in
troducing an improved version of SDS namely variational
score distillation VSD to augment and finetune a pre
trained diffusion model and use it for generation
Diffusion DistillationBased 3D Scene Editing Similar
to 28 for instructionguided generation tasks another dif
fusion model 2 was proposed for instructionguided im
age editing by generating the edited image conditioned on
both the original image and the instruction which is there
fore compatible with SDS 25 Instruction 3Dto3D 13
uses SDS with 2 to support instructionguided style trans
fer on 3D scenes InstructNeRF2NeRF IN2N 8 adopts
another way to operate the 2D diffusion model similar to
the rephrased version of SDS in HiFA 48 which iteratively generates edited images to update the NeRF dataset
for NeRF fitting and supports more general editing instruc
tions such as objectspecific editing ViCANeRF 6 pro
poses a different pipeline to first edit key views and then
blend key views and apply refinement EditDiffNeRF 45
augments the diffusion model and finetunes it with a CLIP
loss to improve the success rate of editing DreamEdi
tor 49 utilizes a finetuned variant of 29 instead of 2
and focuses on objectspecific editing
Consistency in DistillationBased Pipelines When dis
tilling from 2D diffusion to perform 3D generation or edit
ing 3D awareness and 3D consistency of the generated
images are crucial as 3Dinconsistent multiview images
are not a valid descriptor of a scene However achiev
ing 3D consistency in 2D diffusion is challenging Early
work 8 13 18 25 48 does not alter the diffusion model
and relies on consistency derived from NeRF by directly
training NeRF with the inconsistent multiview images The
NeRF will then converge to an averaged or smoothed ver
sion of the scene according to its model capability which
results in blurred results with few textures and even fails
to generate regular patterns like a plaid or checkered pat
tern Followup work begins to improve the consistency of
the diffusion andor the pipeline ViCANeRF 6 achieves
consistency by proposing a different pipeline based on key
views ProlificDreamer 36 makes the diffusion 3Daware
by inputting the camera parameter to the diffusion model
and applying perscene finetuning CSD 15 IVID 37
and ConsistNet 42 propose a joint distillation procedure
for multiple views aiming to generate or edit multiple im
ages in one batch consistently through either attention
depthbased warping or KullbackLeibler divergence
However these methods all share two major constraints
1 the noise used for generation is not controlled therefore
a single view may lead to different and inconsistent gener
ation results with different noises 2 these methods only
study and enforce the consistency between images within
a single batch Nevertheless the full generation or editing
procedure for the scene is across multiple batches and there
might be inconsistencies in different batches Our Con
sistDreamer resolves these limitations by proposing novel
structured noise and consistencyenforcing training
3 ConsistDreamer Methodology
Our ConsistDreamer is a novel IN2Nlike 8 framework

manipulating noise 3We evaluate a range of scenes and

editing instructions achieving stateoftheart performance

in both scenes considered by previous work and more com

plicated largescale indoor scenes from ScanNet

22 Related Work

NeRFBased Scene Editing Neural radiance field

NeRF 21 and its variants 1 3 7 19 33 35 38 44

are widelyused approaches to representing scenes NeRF

leverages neural networks or other learnable architectures to

learn to reconstruct the 3D geometry of a scene only from

multiview images and their camera parameters and sup

port novel view synthesis With the development of NeRF

editing a NeRFrepresented scene is also deeply studied

covering different types of editing objectives and editing

operation indicators aka user interfaces Some meth

ods 16 20 41 support editing the position color andor

shape of a specific object indicated by users through a

pixel a text description or a segment etc Another line

of work 4 24 39 46 studies humanguided shape edit

ing which allows users to indicate a shape editing oper

ation with a cage or point cloud provided by the model

The task we investigate is instructionguided scene editing

which allows users to indicate the editing operation through

instructions in natural language The first work in this di

rection is NeRFArt 34 which mainly focuses on style

transfer instructions and uses pretrained CLIP 26 as the

stylization loss for the instructionindicated style More re

cent work 8 13 15 45 49 leverages diffusion models 10

instead of CLIP to benefit from powerful diffusion models

and support more general instructions

DistillationBased 3D Scene Generation Lacking 3D

datasets to train powerful 3D diffusion models current so

lutions distill the generation signal from a 2D diffusion

model to exploit its ability in 3D generation DreamFu

sion 25 is the first work in this direction which proposes

score distillation sampling SDS to distill the gradient up

date direction score from 2D diffusion models and

supports instructionguided scene generation by distilling a

pretrained diffusion model 28 HiFA 48 proposes an an

nealing technique and rephrases the distillation formula to

improve the generation result Magic3D 18 improves the

generation results by introducing a coarsetofine strategy

and a mesh exportation and refinement method Prolific

Dreamer 36 further improves the generation results by in

troducing an improved version of SDS namely variational

score distillation VSD to augment and finetune a pre

trained diffusion model and use it for generation

Diffusion DistillationBased 3D Scene Editing Similar

to 28 for instructionguided generation tasks another dif

fusion model 2 was proposed for instructionguided im

age editing by generating the edited image conditioned on

both the original image and the instruction which is there

fore compatible with SDS 25 Instruction 3Dto3D 13

uses SDS with 2 to support instructionguided style trans

fer on 3D scenes InstructNeRF2NeRF IN2N 8 adopts

another way to operate the 2D diffusion model similar to

the rephrased version of SDS in HiFA 48 which iteratively generates edited images to update the NeRF dataset

for NeRF fitting and supports more general editing instruc

tions such as objectspecific editing ViCANeRF 6 pro

poses a different pipeline to first edit key views and then

blend key views and apply refinement EditDiffNeRF 45

augments the diffusion model and finetunes it with a CLIP

loss to improve the success rate of editing DreamEdi

tor 49 utilizes a finetuned variant of 29 instead of 2

and focuses on objectspecific editing

Consistency in DistillationBased Pipelines When dis

tilling from 2D diffusion to perform 3D generation or edit

ing 3D awareness and 3D consistency of the generated

images are crucial as 3Dinconsistent multiview images

are not a valid descriptor of a scene However achiev

ing 3D consistency in 2D diffusion is challenging Early

work 8 13 18 25 48 does not alter the diffusion model

and relies on consistency derived from NeRF by directly

training NeRF with the inconsistent multiview images The

NeRF will then converge to an averaged or smoothed ver

sion of the scene according to its model capability which

results in blurred results with few textures and even fails

to generate regular patterns like a plaid or checkered pat

tern Followup work begins to improve the consistency of

the diffusion andor the pipeline ViCANeRF 6 achieves

consistency by proposing a different pipeline based on key

views ProlificDreamer 36 makes the diffusion 3Daware

by inputting the camera parameter to the diffusion model

and applying perscene finetuning CSD 15 IVID 37

and ConsistNet 42 propose a joint distillation procedure

for multiple views aiming to generate or edit multiple im

ages in one batch consistently through either attention

depthbased warping or KullbackLeibler divergence

However these methods all share two major constraints

1 the noise used for generation is not controlled therefore

a single view may lead to different and inconsistent gener

ation results with different noises 2 these methods only

study and enforce the consistency between images within

a single batch Nevertheless the full generation or editing

procedure for the scene is across multiple batches and there

might be inconsistencies in different batches Our Con

sistDreamer resolves these limitations by proposing novel

structured noise and consistencyenforcing training

3 ConsistDreamer Methodology

Our ConsistDreamer is a novel IN2Nlike 8 framework