Yinhuai Wang*, Jiwen Yu*, Jian Zhang
Peking University and PCL
*denotes equal contribution
This repository contains the code release for Zero Shot Image Restoration Using Denoising Diffusion Null-Space Model. DDNM can solve various image restoration tasks without any optimization or training! Yes, in a zero-shot manner.
Supported Applications:
- Arbitrary Size🆕
- Old Photo Restoration🆕
- Super-Resolution
- Denoising
- Colorization
- Inpainting
- Deblurring
- Compressed Sensing
- Add other tasks to the HQ version.
git clone https://github.com/wyhuai/DDNM.git
pip install numpy torch blobfile tqdm pyYaml pillow # e.g. torch 1.7.1+cu110.
To restore human face images, download this model(from SDEdit) and put it into DDNM/exp/logs/celeba/
.
wget https://image-editing-test-12345.s3-us-west-2.amazonaws.com/checkpoints/celeba_hq.ckpt
To restore general images, download this model(from guided-diffusion) and put it into DDNM/exp/logs/imagenet/
.
wget https://openaipublic.blob.core.windows.net/diffusion/jul-2021/256x256_diffusion_uncond.pt
Run below command to get 4x SR results immediately. The results should be in DDNM/exp/image_samples/demo
.
python main.py --ni --simplified --config celeba_hq.yml --path_y celeba_hq --eta 0.85 --deg "sr_averagepooling" --deg_scale 4.0 --sigma_y 0 -i demo
The detailed sampling command is here:
python main.py --ni --simplified --config {CONFIG}.yml --path_y {PATH_Y} --eta {ETA} --deg {DEGRADATION} --deg_scale {DEGRADATION_SCALE} --sigma_y {SIGMA_Y} -i {IMAGE_FOLDER}
with following options:
- We implement TWO versions of DDNM in this repository. One is SVD-based version, which is more precise in solving noisy tasks. Another one is the simplified version, which does not involve SVD and is flexible for users to define their own degradations. Use
--simplified
to activate the simplified DDNM. Without--simplified
will turn to the SVD-based DDNM. PATH_Y
is the folder name of the test dataset, inDDNM/exp/datasets
.ETA
is the DDIM hyperparameter. (default:0.85
)DEGREDATION
is the supported tasks includingcs_walshhadamard
,cs_blockbased
,inpainting
,denoising
,deblur_uni
,deblur_gauss
,deblur_aniso
,sr_averagepooling
,sr_bicubic
,colorization
,mask_color_sr
, and user-defineddiy
.DEGRADATION_SCALE
is the scale of degredation. e.g.,--deg sr_bicubic --deg_scale 4
lead to 4xSR.SIGMA_Y
is the noise observed in y.CONFIG
is the name of the config file (seeconfigs/
for a list), including hyperparameters such as batch size and sampling step.IMAGE_FOLDER
is the folder name of the results.
For the config files, e.g., celeba_hq.yml, you may change following properties:
sampling:
batch_size: 1
time_travel:
T_sampling: 100 # sampling steps
travel_length: 1 # time-travel parameters l and s, see section 3.3 of the paper.
travel_repeat: 1 # time-travel parameter r, see section 3.3 of the paper.
Dataset download link: [Google drive] [Baidu drive]
Download the CelebA testset and put it into DDNM/exp/datasets/celeba/
.
Download the ImageNet testset and put it into DDNM/exp/datasets/imagenet/
and replace the file DDNM/exp/imagenet_val_1k.txt
.
Run the following command. You may increase the batch_size to accelerate evaluation.
sh evaluation.sh
The High-Quality results presented in the front figure are mostly generated by applying DDNM to the models in RePaint, which uses time-travel trick based on DDPM sampling (see the part "DDNM for Arbitrary Size"). By the way, we find that using DDIM (without time-travel) usually yields better results than DDPM (without time-travel). However, DDPM (with time-travel) yields better results than DDIM (with time-travel).
Run the following command
python main.py --ni --simplified --config celeba_hq.yml --path_y solvay --eta 0.85 --deg "sr_averagepooling" --deg_scale 4.0 --sigma_y 0.1 -i demo
Run the following command
python main.py --ni --simplified --config oldphoto.yml --path_y oldphoto --eta 0.85 --deg "mask_color_sr" --deg_scale 2.0 --sigma_y 0.02 -i demo
You may use DDNM to restore your own degraded images. DDNM provides full flexibility for you to define the degradation operator and the noise level. Note that these definitions are critical for a good results. You may reference the following guidance.
- If your are using CelebA pretrained models, try this tool to crop and align your photo.
- If there are local artifacts on your photo, try this tool to draw a mask to cover them, and save this mask to
DDNM/exp/inp_masks/mask.png
. Then runDDNM/exp/inp_masks/get_mask.py
to generatemask.npy
. - If your photo is faded, you need a grayscale operator as part of the degradation.
- If your photo is blur, you need a downsampler operator as part of the degradation. Also, you need to set a proper SR scale
--deg_scale
. - If your photo suffers global artifacts, e.g., jpeg-like artifacts or unkown noise, you need to set a proper
sigma_y
to remove these artifacts. - Search
args.deg =='diy'
inDDNM/guided_diffusion/diffusion.py
and change the definition of$\mathbf{A}$ correspondingly. Then run
python main.py --ni --simplified --config celeba_hq.yml --path_y {YOUR_OWN_PATH} --eta 0.85 --deg "diy" --deg_scale {YOUR_OWN_SCALE} --sigma_y {YOUR_OWN_LEVEL} -i diy
Above we show an example of using DDNM to SR a 64x256 input image into a 256x1024 result. We call this newly conceived method the Mask-Shift trick, whose details can be found in the updated version of our paper, section 7.
We implement the Mask-Shift trick in the folder hq_demo
, based on RePaint. To try this demo, you need to download the pre-trained models:
wget https://openaipublic.blob.core.windows.net/diffusion/jul-2021/256x256_classifier.pt
wget https://openaipublic.blob.core.windows.net/diffusion/jul-2021/256x256_diffusion.pt
and put it to hq_demo/data/pretrained
. Then run
cd hq_demo
sh evaluation.sh
This script contains SR results up to 2K resolution. It may take hours to finish some demos in this script. Setting a smaller sampling step or time-travel parameters in hq_demo/confs/inet256.yml can speed up, but may compromise the generative quality.
It is very easy to implement a basic DDNM on your own diffusion model! You may reference the following:
- Copy these operator implementations to the core diffusion sampling file, then define your task type, e.g., set
IR_mode="super resolution"
.
def color2gray(x):
coef=1/3
x = x[:,0,:,:] * coef + x[:,1,:,:]*coef + x[:,2,:,:]*coef
return x.repeat(1,3,1,1)
def gray2color(x):
x = x[:,0,:,:]
coef=1/3
base = coef**2 + coef**2 + coef**2
return th.stack((x*coef/base, x*coef/base, x*coef/base), 1)
def PatchUpsample(x, scale):
n, c, h, w = x.shape
x = torch.zeros(n,c,h,scale,w,scale) + x.view(n,c,h,1,w,1)
return x.view(n,c,scale*h,scale*w)
# Implementation of A and its pseudo-inverse Ap
if IR_mode=="colorization":
A = color2gray
Ap = gray2color
elif IR_mode=="inpainting":
A = lambda z: z*mask
Ap = A
elif IR_mode=="super resolution":
A = torch.nn.AdaptiveAvgPool2d((256//scale,256//scale))
Ap = lambda z: PatchUpsample(z, scale)
elif IR_mode=="old photo restoration":
A1 = lambda z: z*mask
A1p = A1
A2 = color2gray
A2p = gray2color
A3 = torch.nn.AdaptiveAvgPool2d((256//scale,256//scale))
A3p = lambda z: PatchUpsample(z, scale)
A = lambda z: A3(A2(A1(z)))
Ap = lambda z: A1p(A2p(A3p(z)))
- Find the variant
$\mathbf{x}_{0|t}$ in the target codes, and use the result of this function to modify the sampling of$\mathbf{x}_{t-1}$ . Your may need to provide the input degraded image$\mathbf{y}$ and the corresponding noise level$\sigma_\mathbf{y}$ .
# Core Implementation of DDNM+, simplified denoising solution (Section 3.3).
# For more accurate denoising, please refer to the paper (Appendix I) and the source code.
def ddnm_plus_core(x0t, y, sigma_y=0, sigma_t, a_t):
#Eq 19
if sigma_t >= a_t*sigma_y:
lambda_t = 1
gamma_t = sigma_t**2 - (a_t*lambda_t*sigma_y)**2
else:
lambda_t = sigma_t/(a_t*sigma_y)
gamma_t = 0
#Eq 17
x0t= x0t + lambda_t*Ap(y - A(x0t))
return x0t, gamma_t
- Actually, this repository contains the above simplified implementation: try search
arg.simplified
inDDNM/guided_diffusion/diffusion.py
for related codes.
If you find this repository useful for your research, please cite the following work.
@article{wang2022ddnm,
title={Zero Shot Image Restoration Using Denoising Diffusion Null-Space Model},
author={Yinhuai, Wang and Jiwen, Yu and Jian, Zhang},
journal={arXiv:2212.00490},
year={2022}}
}
This implementation is based on / inspired by:
- https://github.com/wyhuai/RND (null-space learning)
- https://github.com/andreas128/RePaint (time-travel trick)
- https://github.com/bahjat-kawar/ddrm (code structure)