tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
crossbatch consistency for views edited by different GPUs
Consistent Denoising Procedure With our structured
noise the denoising in 2D diffusion initiates with consis
tent noise This leads to a further goal to make the entire
denoising procedure 3D consistent and thus end with con
sistent images We achieve this by enforcing all the views in
the intermediate denoising images to be also 3D consistent
at each denoising step Therefore unlike the conventional
diffusion training with singlestep denoising our training
involves a full multistep denoising procedure with passing
through gradients As it is impossible to fit the entire com
putational graph into the GPU memory we use checkpoint
ing 27 40 to trade space with time Doing so enables con
structing the reference set of images with warping for each
intermediate denoising step which is then used to super
vise the intermediate denoising image This provides more
direct signals of 3D consistency in the training of diffusionfacilitating the generation of 3D consistent results
Shape Editing Some instructions egMake him smile
change the shape or geometry of the scene during editing
while our structured noise and consistencyenforcing train
ing rely on the geometry To be compatible with shape
editing we design a coarsetofine strategy we first edit
the scene using ConsistDreamer with only the surround
ing view and disabling the other two components ie us
ing imageindependent noise and the original implementa
tion of 2 This allows the scene to converge to a coarse
edited shape according to the instruction We then activate
structured noise and consistencyenforcing training to re
fine the editing We periodically adjust the structured noise
with changes in geometry while preserving the noise val
ues With this strategy ConsistDreamer also achieves high
fidelity shape editing
4 Experiments
Editing Tasks In our setting each editing task is a pair of
scene instruction indicating which instruction
guided editing operation should be applied on which scene
The output of the task is another scene being the edited
scene under the instruction The scenes we use for evalua
tion contain two parts 1 IN2N Scenes used by IN2N 21
including scenes of human faces or bodies outdoor scenes
and statues and 2 SN Scenes in ScanNet 43
which are complicated indoor scenes with freeformed
structures and camera trajectories We also use two types
of editing instructions 1 style transfer which transfers the
style of the scene into the described style and 2 object
specific editing which edits a specific object of the scene
6We use these tasks to compare our approach with baselines
and conduct ablation study on representative tasks
NeRF Backbone and Diffusion Model For a fair com
parison with previous works 6 8 we use the Nerfacto
model in NeRFStudio 32 as our NeRF backbone and the
pretrained diffusion model 2 from Hugging Face as our
initial checkpoint The NeRF representation for the scene is
trained with NeRFStudio in advance and then used in our
pipeline
ConsistDreamer Variants We investigate the following
variants for our ablation study where SNSV and T
denote removing structured noise surrounding views and
consistencyenforcing training respectively 1 Full Con
sistDreamer 2 No structured noise SN use inde
pendently generated noise for each view instead of struc
tured noise but still use surrounding views and perform
consistencyenforcing training 3 No training T use
surrounding views and structured noise but do not aug
ment and train 2 and keep using the original checkpoint
4 Only surrounding views SNT only use surround
ing views and do not use structured noise or train 2
5 IN2N SNSVT ours with all the proposed
components removed which can be regarded as an alterna
tive version of IN2N Note that consistencyenforcing train
ing requires surrounding views to produce sufficient edited
views in one generation we cannot remove surrounding
views but still apply consistencyenforcing training on 2
Baselines We mainly compare our method with two
baselines InstructNeRF2NeRF IN2N 8 and ViCA
NeRF ViCA 6 as they are most closely related to our
task We also compare with NeRFArt NArt 34 as an
early work Other methods however lack publicly avail
able or working code andor only use a few scenes sup
ported by NerfStudio Therefore we could only compare
with CSD 15 DreamEditor 49 GE 5 EN2N 31
and PDS 17 under a few tasks in supplementary and
are unable to compare with EditDiffNeRF 45 and In
struct 3Dto3D 13 Note that ConsistDreamer solves
instructionguided scene editing instead of scene genera
tion so we do not compare with models for the generation
task 25 37 42 48
Evaluation Metrics Observing that our ConsistDreamer
generates significantly sharper editing results consistent
with previous work 6 8 we compare ConsistDreamer
with baselines mainly through qualitative evaluation For
the ablation study the appearance of the scenes edited by
our different variants may be visually similar and unable
to be fairly compared using qualitative results Therefore
we propose distillation fidelity score DFS to evaluate how
faithful the editing is distilled and applied on NeRF com
pared with the diffusions output 2 rooted in the basic
setting that we distill from 2 to edit 3D scenes In this
situation our editing ability is bounded by 2s Consistent
Variant Components A B C D

crossbatch consistency for views edited by different GPUs

Consistent Denoising Procedure With our structured

noise the denoising in 2D diffusion initiates with consis

tent noise This leads to a further goal to make the entire

denoising procedure 3D consistent and thus end with con

sistent images We achieve this by enforcing all the views in

the intermediate denoising images to be also 3D consistent

at each denoising step Therefore unlike the conventional

diffusion training with singlestep denoising our training

involves a full multistep denoising procedure with passing

through gradients As it is impossible to fit the entire com

putational graph into the GPU memory we use checkpoint

ing 27 40 to trade space with time Doing so enables con

structing the reference set of images with warping for each

intermediate denoising step which is then used to super

vise the intermediate denoising image This provides more

direct signals of 3D consistency in the training of diffusionfacilitating the generation of 3D consistent results

Shape Editing Some instructions egMake him smile

change the shape or geometry of the scene during editing

while our structured noise and consistencyenforcing train

ing rely on the geometry To be compatible with shape

editing we design a coarsetofine strategy we first edit

the scene using ConsistDreamer with only the surround

ing view and disabling the other two components ie us

ing imageindependent noise and the original implementa

tion of 2 This allows the scene to converge to a coarse

edited shape according to the instruction We then activate

structured noise and consistencyenforcing training to re

fine the editing We periodically adjust the structured noise

with changes in geometry while preserving the noise val

ues With this strategy ConsistDreamer also achieves high

fidelity shape editing

4 Experiments

Editing Tasks In our setting each editing task is a pair of

scene instruction indicating which instruction

guided editing operation should be applied on which scene

The output of the task is another scene being the edited

scene under the instruction The scenes we use for evalua

tion contain two parts 1 IN2N Scenes used by IN2N 21

including scenes of human faces or bodies outdoor scenes

and statues and 2 SN Scenes in ScanNet 43

which are complicated indoor scenes with freeformed

structures and camera trajectories We also use two types

of editing instructions 1 style transfer which transfers the

style of the scene into the described style and 2 object

specific editing which edits a specific object of the scene

6We use these tasks to compare our approach with baselines

and conduct ablation study on representative tasks

NeRF Backbone and Diffusion Model For a fair com

parison with previous works 6 8 we use the Nerfacto

model in NeRFStudio 32 as our NeRF backbone and the

pretrained diffusion model 2 from Hugging Face as our

initial checkpoint The NeRF representation for the scene is

trained with NeRFStudio in advance and then used in our

pipeline

ConsistDreamer Variants We investigate the following

variants for our ablation study where SNSV and T

denote removing structured noise surrounding views and

consistencyenforcing training respectively 1 Full Con

sistDreamer 2 No structured noise SN use inde

pendently generated noise for each view instead of struc

tured noise but still use surrounding views and perform

consistencyenforcing training 3 No training T use

surrounding views and structured noise but do not aug

ment and train 2 and keep using the original checkpoint

4 Only surrounding views SNT only use surround

ing views and do not use structured noise or train 2

5 IN2N SNSVT ours with all the proposed

components removed which can be regarded as an alterna

tive version of IN2N Note that consistencyenforcing train

ing requires surrounding views to produce sufficient edited

views in one generation we cannot remove surrounding

views but still apply consistencyenforcing training on 2

Baselines We mainly compare our method with two

baselines InstructNeRF2NeRF IN2N 8 and ViCA

NeRF ViCA 6 as they are most closely related to our

task We also compare with NeRFArt NArt 34 as an

early work Other methods however lack publicly avail

able or working code andor only use a few scenes sup

ported by NerfStudio Therefore we could only compare

with CSD 15 DreamEditor 49 GE 5 EN2N 31

and PDS 17 under a few tasks in supplementary and

are unable to compare with EditDiffNeRF 45 and In

struct 3Dto3D 13 Note that ConsistDreamer solves

instructionguided scene editing instead of scene genera

tion so we do not compare with models for the generation

task 25 37 42 48

Evaluation Metrics Observing that our ConsistDreamer

generates significantly sharper editing results consistent

with previous work 6 8 we compare ConsistDreamer

with baselines mainly through qualitative evaluation For

the ablation study the appearance of the scenes edited by

our different variants may be visually similar and unable

to be fairly compared using qualitative results Therefore

we propose distillation fidelity score DFS to evaluate how

faithful the editing is distilled and applied on NeRF com

pared with the diffusions output 2 rooted in the basic

setting that we distill from 2 to edit 3D scenes In this

situation our editing ability is bounded by 2s Consistent

Variant Components A B C D