tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
D1 Hyperparameters and Settings
In our experiments we use the multiGPU pipeline with
n 3 4 GPUs in total and surrounding views of k 5
1 main view and 12 reference views
The learning rate of each component is shown below
NeRFacto 32 5103for field part and 102for
proposal network part
LoRAaugmented diffusion model 2 consistent with
their original implementation 104
Learnable 3D positional embedding 2103
All the views are resized to 3 4 or4 3 according
to their orientations For landscape images portrait images
use the same setting with a flipped height and width the
diffusion model takes a surrounding view image input at
1152864also 4 3 with horizontal splitters at heights
6pix and vertical splitters at widths 8pix The sizes of the
main view and reference views are 688516and224
168 respectively This setting is consistent with the original
usage of diffusion models 2 28 trained at 512512 as
our main view has a height close to it
Consistent with IN2N 8 both MSE and LPIPS losses
are used to train NeRF
12Figure B4 Our ConsistDreamer outperforms all three concurrent works with brighter color clearer textures and better editing results
matched with the editing instruction
MethodTextImage Direction Similarity CTIDS Direction Consistency CDC
A B C D A B C D
Ours 00259 01679 01204 01268 05785 01735 03077 02878
IN2N 8 00099 01252 01163 01055 05106 01634 02900 01772
Table C1 Our ConsistDreamer significantly and consistently outperforms baseline IN2N in CLIP 26 metrics over all four ablation
scenes
Figure D1 Our multiGPU training pipeline uses n 1GPUs one dedicated for NeRF fitting and the others for diffusion training and
generation
D2 Viewpoints And Camera Trajectory
During the distillation process we directly use the view
points provided in the original scene dataset which is suf
ficient to cover the whole scene In visualization we use
the provided camera trajectory for IN2N 8 dataset and
manually construct another camera trajectory for a smooth
visualization for ScanNet 43 datasetD3 Training Schedule
One standard full training contains 1600 epochs across
multiple substages All these stages are explained below
Initialization Stage Epoch 1200 Train diffusion 2
before NeRF fitting We perform one diffusion training
step in one epoch
Early Bootstrap Epoch 150 Train the LoRA
augmented diffusion model to mimic the behavior of
the original model with the augmented input of 3D po
13sitional embedding The weight regularization loss of
maintaining original behavior detailed in D8 is sig
nificantly higher NeRF training has not started
Bootstrap Epoch 51150 Train the consistency
awareness of the LoRAaugmented diffusion model
while keeping original behavior at a similar impor
tance with balanced weights
Warming Up Epoch 151200 Use the standard
weights to balance the consistency loss and regulariza
tion focusing more on consistency This epoch gen
erates sufficient images for the edited view buffer for
NeRF fitting
Distillation Stage Epoch 2011600 Train diffusion
while fitting NeRF In each of 4 epochs we do 3 diffu
sion generation steps without training to fill the edited
view buffer and only one diffusion training step Here
the noise level means the mixture rate of the current
NeRF being edited rendered image and the noise as the
diffusions input full noise level means using only noise
for generation standard generation while a 30 noise
level means the input image is the mixture of 30 noise
and70 rendered image
Full Noise Generation Epoch 201500 The diffu
sion model is trained and used for generation at a full
noise level to edit the views sufficiently regardless of
NeRF
PreAnnealing Epoch 501600 The diffusion
model is trained and used for generation with a noise
level sampled from 70100 It edits the views
with a few references to the current NeRF starting to
refine the current NeRF
Annealing Epoch 6011500 Following the idea
of HiFA 48 the range of the noise level linearly an
neals from 70100 to1040 The NeRF will
gradually converge to a finegrained edited version
Ending Epoch 15011600 The diffusion model is
trained and used for generation with a noise level sam
pled from the annealed range 1040 to further
refine the edited NeRF
If the editing task requires editing the geometry or shape
of the scene shape editing the depthbased warping us
ing the depth of the original scene will be inaccurate There
fore in the Initialization Stage we put the original diffu
sion models output to the edited view buffer for NeRF fit
ting equivalently using IN2N in this stage In the distilla
tion stage the shape of the NeRF will be adjusted to the
edited shape in a short time and then we will start to use
the trained diffusion models output for NeRF fitting
In our experiments most IN2N scenes converge to a fine
grained edited scene at 600700 epochs while Scan
Net 43 scenes take around 1000 epochsD4 Structured Noise Implementation
In the main paper the structured noise is implemented by
constructing a dense point cloud of the scene by unproject

D1 Hyperparameters and Settings

In our experiments we use the multiGPU pipeline with

n 3 4 GPUs in total and surrounding views of k 5

1 main view and 12 reference views

The learning rate of each component is shown below

NeRFacto 32 5103for field part and 102for

proposal network part

LoRAaugmented diffusion model 2 consistent with

their original implementation 104

Learnable 3D positional embedding 2103

All the views are resized to 3 4 or4 3 according

to their orientations For landscape images portrait images

use the same setting with a flipped height and width the

diffusion model takes a surrounding view image input at

1152864also 4 3 with horizontal splitters at heights

6pix and vertical splitters at widths 8pix The sizes of the

main view and reference views are 688516and224

168 respectively This setting is consistent with the original

usage of diffusion models 2 28 trained at 512512 as

our main view has a height close to it

Consistent with IN2N 8 both MSE and LPIPS losses

are used to train NeRF

12Figure B4 Our ConsistDreamer outperforms all three concurrent works with brighter color clearer textures and better editing results

matched with the editing instruction

MethodTextImage Direction Similarity CTIDS Direction Consistency CDC

A B C D A B C D

Ours 00259 01679 01204 01268 05785 01735 03077 02878

IN2N 8 00099 01252 01163 01055 05106 01634 02900 01772

Table C1 Our ConsistDreamer significantly and consistently outperforms baseline IN2N in CLIP 26 metrics over all four ablation

scenes

Figure D1 Our multiGPU training pipeline uses n 1GPUs one dedicated for NeRF fitting and the others for diffusion training and

generation

D2 Viewpoints And Camera Trajectory

During the distillation process we directly use the view

points provided in the original scene dataset which is suf

ficient to cover the whole scene In visualization we use

the provided camera trajectory for IN2N 8 dataset and

manually construct another camera trajectory for a smooth

visualization for ScanNet 43 datasetD3 Training Schedule

One standard full training contains 1600 epochs across

multiple substages All these stages are explained below

Initialization Stage Epoch 1200 Train diffusion 2

before NeRF fitting We perform one diffusion training

step in one epoch

Early Bootstrap Epoch 150 Train the LoRA

augmented diffusion model to mimic the behavior of

the original model with the augmented input of 3D po

13sitional embedding The weight regularization loss of

maintaining original behavior detailed in D8 is sig

nificantly higher NeRF training has not started

Bootstrap Epoch 51150 Train the consistency

awareness of the LoRAaugmented diffusion model

while keeping original behavior at a similar impor

tance with balanced weights

Warming Up Epoch 151200 Use the standard

weights to balance the consistency loss and regulariza

tion focusing more on consistency This epoch gen

erates sufficient images for the edited view buffer for

NeRF fitting

Distillation Stage Epoch 2011600 Train diffusion

while fitting NeRF In each of 4 epochs we do 3 diffu

sion generation steps without training to fill the edited

view buffer and only one diffusion training step Here

the noise level means the mixture rate of the current

NeRF being edited rendered image and the noise as the

diffusions input full noise level means using only noise

for generation standard generation while a 30 noise

level means the input image is the mixture of 30 noise

and70 rendered image

Full Noise Generation Epoch 201500 The diffu

sion model is trained and used for generation at a full

noise level to edit the views sufficiently regardless of

NeRF

PreAnnealing Epoch 501600 The diffusion

model is trained and used for generation with a noise

level sampled from 70100 It edits the views

with a few references to the current NeRF starting to

refine the current NeRF

Annealing Epoch 6011500 Following the idea

of HiFA 48 the range of the noise level linearly an

neals from 70100 to1040 The NeRF will

gradually converge to a finegrained edited version

Ending Epoch 15011600 The diffusion model is

trained and used for generation with a noise level sam

pled from the annealed range 1040 to further

refine the edited NeRF

If the editing task requires editing the geometry or shape

of the scene shape editing the depthbased warping us

ing the depth of the original scene will be inaccurate There

fore in the Initialization Stage we put the original diffu

sion models output to the edited view buffer for NeRF fit

ting equivalently using IN2N in this stage In the distilla

tion stage the shape of the NeRF will be adjusted to the

edited shape in a short time and then we will start to use

the trained diffusion models output for NeRF fitting

In our experiments most IN2N scenes converge to a fine

grained edited scene at 600700 epochs while Scan

Net 43 scenes take around 1000 epochsD4 Structured Noise Implementation

In the main paper the structured noise is implemented by

constructing a dense point cloud of the scene by unproject