Awesome

DiffBIR: Towards Blind Image Restoration with Generative Diffusion Prior

Paper | Project Page

Xinqi Lin^1,*, Jingwen He^2,3,*, Ziyan Chen¹, Zhaoyang Lyu², Bo Dai², Fanghua Yu¹, Wanli Ouyang², Yu Qiao², Chao Dong^1,2

¹Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences<br>²Shanghai AI Laboratory<br>³The Chinese University of Hong Kong

:star:If DiffBIR is helpful for you, please help star this repo. Thanks!:hugs:

:book:Table Of Contents

• Update
• Visual Results On Real-world Images
• TODO
• Installation
• Quick Start
• Pretrained Models
• Inference
• Train

<a name="update"></a>:new:Update

• 2024.11.27: ✅ Release DiffBIR v2.1, including a new model trained on unsplash dataset with LLaVA-generated captions, more samplers, better tiled-sampling support and so on. Check release note for details.
• 2024.04.08: ✅ Release everything about our updated manuscript, including (1) a new model trained on subset of laion2b-en and (2) a more readable code base, etc. DiffBIR is now a general restoration pipeline that could handle different blind image restoration tasks with a unified generation module.
• 2023.09.19: ✅ Add support for Apple Silicon! Check installation_xOS.md to work with CPU/CUDA/MPS device!
• 2023.09.14: ✅ Integrate a patch-based sampling strategy (mixture-of-diffusers). Try it! Here is an example with a resolution of 2396 x 1596. GPU memory usage will continue to be optimized in the future and we are looking forward to your pull requests!
• 2023.09.14: ✅ Add support for background upsampler (DiffBIR/RealESRGAN) in face enhancement! :rocket: Try it!
• 2023.09.13: :rocket: Provide online demo (DiffBIR-official) in OpenXLab, which integrates both general model and face model. Please have a try! camenduru also implements an online demo, thanks for his work.:hugs:
• 2023.09.12: ✅ Upload inference code of latent image guidance and release real47 testset.
• 2023.09.08: ✅ Add support for restoring unaligned faces.
• 2023.09.06: :rocket: Update colab demo. Thanks to camenduru!:hugs:
• 2023.08.30: This repo is released.

<a name="visual_results"></a>:eyes:Visual Results On Real-world Images

Blind Image Super-Resolution

<img src="assets/visual_results/bsr6.png" height="223px"/> <img src="assets/visual_results/bsr7.png" height="223px"/> <img src="assets/visual_results/bsr4.png" height="223px"/>

Blind Face Restoration

<img src="assets/visual_results/whole_image1.png" height="370"/> <img src="assets/visual_results/whole_image2.png" height="370"/>

:star: Face and the background enhanced by DiffBIR.

Blind Image Denoising

<img src="assets/visual_results/bid1.png" height="215px"/> <img src="assets/visual_results/bid3.png" height="215px"/> <img src="assets/visual_results/bid2.png" height="215px"/>

8x Blind Super-Resolution With Tiled Sampling

I often think of Bag End. I miss my books and my arm chair, and my garden. See, that's where I belong. That's home. --- Bilbo Baggins

<img src="assets/visual_results/tiled_sampling.png" height="480px"/>

<a name="todo"></a>:climbing:TODO

• Release code and pretrained models :computer:.
• Update links to paper and project page :link:.
• Release real47 testset :minidisc:.
• Provide webui.
• Reduce the vram usage of DiffBIR :fire::fire::fire:.
• Provide HuggingFace demo :notebook:.
• Add a patch-based sampling schedule :mag:.
• Upload inference code of latent image guidance :page_facing_up:.
• Improve the performance :superhero:.
• Support MPS acceleration for MacOS users.
• DiffBIR-turbo :fire::fire::fire:.
• Speed up inference, such as using fp16/bf16, torch.compile :fire::fire::fire:.

<a name="installation"></a>:gear:Installation

# clone this repo
git clone https://github.com/XPixelGroup/DiffBIR.git
cd DiffBIR

# create environment
conda create -n diffbir python=3.10
conda activate diffbir
pip install -r requirements.txt

Our new code is based on pytorch 2.2.2 for the built-in support of memory-efficient attention. If you are working on a GPU that is not compatible with the latest pytorch, just downgrade pytorch to 1.13.1+cu116 and install xformers 0.0.16 as an alternative.

<a name="quick_start"></a>:flight_departure:Quick Start

Run the following command to interact with the gradio website.

# For low-VRAM users, set captioner to ram or none
python run_gradio.py --captioner llava

<a name="pretrained_models"></a>:dna:Pretrained Models

Here we list pretrained weight of stage 2 model (IRControlNet) and our trained SwinIR, which was used for degradation removal during the training of stage 2 model.

During inference, we use off-the-shelf models from other papers as the stage 1 model: BSRNet for BSR, SwinIR-Face used in DifFace for BFR, and SCUNet-PSNR for BID, while the trained IRControlNet remains unchanged for all tasks. Please check code for more details. Thanks for their work!

<a name="inference"></a>:crossed_swords:Inference

We provide some examples for inference, check inference.py for more arguments. Pretrained weights will be automatically downloaded. For users with limited VRAM, please run the following scripts with tiled sampling.

Blind Image Super-Resolution

# DiffBIR v2 (ECCV paper version)
python -u inference.py \
--task sr \
--upscale 4 \
--version v2 \
--sampler spaced \
--steps 50 \
--captioner none \
--pos_prompt '' \
--neg_prompt 'low quality, blurry, low-resolution, noisy, unsharp, weird textures' \
--cfg_scale 4 \
--input inputs/demo/bsr \
--output results/v2_demo_bsr \
--device cuda --precision fp32

# DiffBIR v2.1
python -u inference.py \
--task sr \
--upscale 4 \
--version v2.1 \
--captioner llava \
--cfg_scale 8 \
--noise_aug 0 \
--input inputs/demo/bsr \
--output results/v2.1_demo_bsr

Blind Aligned-Face Restoration

<a name="inference_fr"></a>

# DiffBIR v2 (ECCV paper version)
python -u inference.py \
--task face \
--upscale 1 \
--version v2 \
--sampler spaced \
--steps 50 \
--captioner none \
--pos_prompt '' \
--neg_prompt 'low quality, blurry, low-resolution, noisy, unsharp, weird textures' \
--cfg_scale 4.0 \
--input inputs/demo/bfr/aligned \
--output results/v2_demo_bfr_aligned \
--device cuda --precision fp32

# DiffBIR v2.1
python -u inference.py \
--task face \
--upscale 1 \
--version v2.1 \
--captioner llava \
--cfg_scale 8 \
--noise_aug 0 \
--input inputs/demo/bfr/aligned \
--output results/v2.1_demo_bfr_aligned

Blind Unaligned-Face Restoration

# DiffBIR v2 (ECCV paper version)
python -u inference.py \
--task face_background \
--upscale 2 \
--version v2 \
--sampler spaced \
--steps 50 \
--captioner none \
--pos_prompt '' \
--neg_prompt 'low quality, blurry, low-resolution, noisy, unsharp, weird textures' \
--cfg_scale 4.0 \
--input inputs/demo/bfr/whole_img \
--output results/v2_demo_bfr_unaligned \
--device cuda --precision fp32

# DiffBIR v2.1
python -u inference.py \
--task face_background \
--upscale 2 \
--version v2.1 \
--captioner llava \
--cfg_scale 8 \
--noise_aug 0 \
--input inputs/demo/bfr/whole_img \
--output results/v2.1_demo_bfr_unaligned

Blind Image Denoising

# DiffBIR v2 (ECCV paper version)
python -u inference.py \
--task denoise \
--upscale 1 \
--version v2 \
--sampler spaced \
--steps 50 \
--captioner none \
--pos_prompt '' \
--neg_prompt 'low quality, blurry, low-resolution, noisy, unsharp, weird textures' \
--cfg_scale 4.0 \
--input inputs/demo/bid \
--output results/v2_demo_bid \
--device cuda --precision fp32

# DiffBIR v2.1
python -u inference.py \
--task denoise \
--upscale 1 \
--version v2.1 \
--captioner llava \
--cfg_scale 8 \
--noise_aug 0 \
--input inputs/demo/bid \
--output results/v2.1_demo_bid

Custom-model Inference

python -u inference.py \
--upscale 4 \
--version custom \
--train_cfg [path/to/training/config] \
--ckpt [path/to/saved/checkpoint] \
--captioner llava \
--cfg_scale 8 \
--noise_aug 0 \
--input inputs/demo/bsr \
--output results/custom_demo_bsr

Other options

Tiled sampling

<a name="patch_based_sampling"></a>

Add the following arguments to enable tiled sampling:

[command...] \
# tiled inference for stage-1 model
--cleaner_tiled \
--cleaner_tile_size 256 \
--cleaner_tile_stride 128 \
# tiled inference for VAE encoding
--vae_encoder_tiled \
--vae_encoder_tile_size 256 \
# tiled inference for VAE decoding
--vae_decoder_tiled \
--vae_decoder_tile_size 256 \
# tiled inference for diffusion process
--cldm_tiled \
--cldm_tile_size 512 \
--cldm_tile_stride 256

Tiled sampling supports super-resolution with a large scale factor on low-VRAM graphics cards. Our tiled sampling is built upon mixture-of-diffusers and Tiled-VAE. Thanks for their work!

Condition as Start Point of Sampling

This option only works with DiffBIR v1 and v2. As proposed in SeeSR, the LR embedding (LRE) strategy provides a more faithful start point for sampling and consequently suppresses the artifacts in flat region:

[command...] --start_point_type cond

For our model, we use the diffused condition as start point. This option makes the results more stable and ensures that the outcomes from ODE samplers like DDIM and DPMS are normal. However, it may lead to a decrease in sample quality.

<a name="train"></a>:stars:Train

Stage 1

First, we train a SwinIR, which will be used for degradation removal during the training of stage 2.

<a name="gen_file_list"></a>

1. Generate file list of training set and validation set, a file list looks like:

   /path/to/image_1
   /path/to/image_2
   /path/to/image_3
   ...
   

   You can write a simple python script or directly use shell command to produce file lists. Here is an example:

   # collect all iamge files in img_dir
   find [img_dir] -type f > files.list
   # shuffle collected files
   shuf files.list > files_shuf.list
   # pick train_size files in the front as training set
   head -n [train_size] files_shuf.list > files_shuf_train.list
   # pick remaining files as validation set
   tail -n +[train_size + 1] files_shuf.list > files_shuf_val.list
   

2. Fill in the training configuration file with appropriate values.

3. Start training!

   accelerate launch train_stage1.py --config configs/train/train_stage1.yaml
   

Stage 2

1. Download pretrained Stable Diffusion v2.1 to provide generative capabilities. :bulb:: If you have ran the inference script, the SD v2.1 checkpoint can be found in weights.

   wget https://huggingface.co/stabilityai/stable-diffusion-2-1-base/resolve/main/v2-1_512-ema-pruned.ckpt --no-check-certificate
   

2. Generate file list as mentioned above. Currently, the training script of stage 2 doesn't support validation set, so you only need to create training file list.

3. Fill in the training configuration file with appropriate values.

4. Start training!

   accelerate launch train_stage2.py --config configs/train/train_stage2.yaml
   

Citation

Please cite us if our work is useful for your research.

@misc{lin2024diffbir,
      title={DiffBIR: Towards Blind Image Restoration with Generative Diffusion Prior}, 
      author={Xinqi Lin and Jingwen He and Ziyan Chen and Zhaoyang Lyu and Bo Dai and Fanghua Yu and Wanli Ouyang and Yu Qiao and Chao Dong},
      year={2024},
      eprint={2308.15070},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

License

This project is released under the Apache 2.0 license.

Acknowledgement

This project is based on ControlNet and BasicSR. Thanks for their awesome work.

Contact

If you have any questions, please feel free to contact with me at linxinqi23@mails.ucas.ac.cn.