AddSR: Accelerating Diffusion-based Blind Super-Resolution with Adversarial Diffusion Distillation

Rui Xie¹ | Ying Tai² | Kai Zhang² | Zhenyu Zhang² | Jun Zhou¹ | Jian Yang³

¹Southwest University, ²Nanjing University, ³Nanjing University of Science and Technology.

:star: If AddSR helps your images or projects, please consider giving this repo a star. Your support is greatly appreciated! :blush:

🔆 Updates

• 2024.04.10 The training code has been released. Please note that it currently only supports a batch size of 2 per device. We will make it possible to support different batchsizes in the future.
• 2024.04.09 The pretrained AddSR model and testing code have been released.

📌 TODO

• ✅ Release the pretrained model
• ✅ Release the training code

🔎 Method Overview

📷 Results Display

⚙️ Dependencies and Installation

## git clone this repository
git clone https://github.com/NJU-PCALab/AddSR.git
cd AddSR

# create an environment with python >= 3.8
conda create -n addsr python=3.8
conda activate addsr
pip install -r requirements.txt

🚀 Inference

Step 1: Download the pretrained models

• Download the pretrained SD-2-base models from HuggingFace.
• Download the AddSR models from GoogleDrive
• Download the DAPE models from GoogleDrive

You can put the models into preset/.

Step 2: Prepare testing data

You can put the testing images in the preset/datasets/test_datasets.

Step 3: Running testing command

python test_addsr.py \
--pretrained_model_path preset/models/stable-diffusion-2-base \
--prompt '' \
--addsr_model_path preset/models/addsr \
--ram_ft_path preset/models/DAPE.pth \
--image_path preset/datasets/test_datasets \
--output_dir preset/datasets/output \
--start_point lr \
--num_inference_steps 4 \
--PSR_weight 0.5

Test Benchmark

RealLR200, RealSR and DRealSR can be downloaded from SeeSR.

🌈 Train

Step1: Download the pretrained models

Download the pretrained SD-2-base models, RAM, SeeSR and DINOv2. You can put them into preset/models.

Step2: Prepare training data

We employ the same preprocessing measures as SeeSR. More details can be found at HERE

Step3: Training for AddSR

 CUDA_VISIBLE_DEVICES="0, 1, 2, 3" accelerate launch train_addsr.py \
--pretrained_model_name_or_path="preset/models/stable-diffusion-2-base" \
--controlnet_model_name_or_path_Tea='preset/seesr' \
--unet_model_name_or_path_Tea='preset/seesr' \
--controlnet_model_name_or_path_Stu='preset/seesr' \
--unet_model_name_or_path_Stu='preset/seesr' \
--dino_path = "preset/models/dinov2_vits14_pretrain.pth" \
--output_dir="./experience/addsr" \
--root_folders 'DataSet/training' \
--ram_ft_path 'preset/models/DAPE.pth' \
--enable_xformers_memory_efficient_attention \
--mixed_precision="fp16" \
--resolution=512 \
--learning_rate=2e-5 \
--train_batch_size=2 \
--gradient_accumulation_steps=2 \
--null_text_ratio=0.5 \
--dataloader_num_workers=0 \
--max_train_steps=50000 \
--checkpointing_steps=5000

• --pretrained_model_name_or_path the path of pretrained SD model from Step 1
• --root_folders the path of your training datasets from Step 2
• --ram_ft_path the path of your DAPE model from Step 3

❤️ Acknowledgments

This project is based on SeeSR, diffusers, BasicSR, ADD and StyleGAN-T. Thanks for their awesome works.

📧 Contact

If you have any inquiries, please don't hesitate to reach out via email at ruixie0097@gmail.com

🎓Citations

If our project helps your research or work, please consider citing our paper:

@misc{xie2024addsr,
      title={AddSR: Accelerating Diffusion-based Blind Super-Resolution with Adversarial Diffusion Distillation}, 
      author={Rui Xie and Ying Tai and Kai Zhang and Zhenyu Zhang and Jun Zhou and Jian Yang},
      year={2024},
      eprint={2404.01717},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

:page_facing_up: License

This project is distributed under the terms of the Apache 2.0 license. Since AddSR is based on SeeSR, StyleGAN-T, and ADD, users must also follow their licenses to use this project.