One-Step Effective Diffusion Network for Real-World Image Super-Resolution

Rongyuan Wu^1,2,*  Lingchen Sun^1,2,*  Zhiyuan Ma^1,*  Lei Zhang^1,2,†

¹The Hong Kong Polytechnic University, ²OPPO Research Institute 

[paper]

---

🚩Accepted by NeurIPS2024

🔥 News

• [2024.10] Updated training codes and paper. 💥💥💥Congratulations, OSEDiff has been applied to the OPPO Find X8 series!
• [2024.07] Release OSEDiff-SD21base.
• [2024.06] This repo is created.

⭐ If OSEDiff is helpful to your images or projects, please help star this repo. Thanks! 🤗

🎬 Overview

🔧 Dependencies and Installation

1. Clone repo

   git clone https://github.com/cswry/OSEDiff.git
   cd OSEDiff

2. Install dependent packages

   conda create -n OSEDiff python=3.10 -y
   conda activate OSEDiff
   pip install --upgrade pip
   pip install -r requirements.txt

3. Download Models

Dependent Models

• SD21 Base
• RAM
• DAPE

⚡ Quick Inference

python test_osediff.py \
-i preset/datasets/test_dataset/input \
-o preset/datasets/test_dataset/output \
--osediff_path preset/models/osediff.pkl \
--pretrained_model_name_or_path SD21BASE_PATH \
--ram_ft_path DAPE_PATH \
--ram_path RAM_PATH

🚀 Calculate Inference Time

For easy comparison, we provide the inference time testing script for osediff.

python test_inference_time.py \
--osediff_path preset/models/osediff.pkl \
--inference_iterations 500 \
--warmup_iterations 5 \
--process_size 512 \
--pretrained_model_name_or_path SD21BASE_PATH \
--ram_ft_path DAPE_PATH  \
--ram_path RAM_PATH

If you run the script on an A100 GPU, you would get

Running 5 warm-up iterations...
Starting inference for 500 iterations...
Inference: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 500/500 [00:52<00:00,  9.55it/s]
Average inference time per iteration: 0.1046 seconds.

📏 Benchmark Results

For convenient comparison, we provide OSEDiff's test results on the RealSR benchmark and DrealSR benchmark in the preset/datasets/benchmark_realsr/results_osediff and preset/datasets/benchmark_drealsr/results_osediff directory. For the DIV2K validation benchmark, due to the large number of images, we have uploaded the test results to Google Drive. These benchmarks were directly copied from StableSR. Additionally, we also provide a script for testing IQA (Image Quality Assessment).

python test_metrics.py \
--inp_imgs preset/datasets/benchmark_realsr/results_osediff \
--gt_imgs preset/datasets/benchmark_realsr/test_HR \
--log preset/datasets/benchmark_realsr/results_osediff/metrics

python test_metrics.py \
--inp_imgs preset/datasets/benchmark_drealsr/results_osediff \
--gt_imgs preset/datasets/benchmark_drealsr/test_HR \
--log preset/datasets/benchmark_drealsr/results_osediff/metrics

If you run these two scripts, you would get them at last line, which are consistent with the data presented in Table 1 of our paper.

RealSR benchmark:
===== Average Metrics for [results_osediff] =====
PSNR: 25.1511; SSIM: 0.7341; LPIPS: 0.2921; DISTS: 0.2128; CLIPIQA: 0.6693; NIQE: 5.6476; MUSIQ: 69.0896; MANIQA: 0.6331 | FID: 123.493817 | FID Runtime: 6.79 sec

DrealSR benchmark:
===== Average Metrics for [results_osediff] =====
PSNR: 27.9243; SSIM: 0.7835; LPIPS: 0.2968; DISTS: 0.2165; CLIPIQA: 0.6963; NIQE: 6.4902; MUSIQ: 64.6537; MANIQA: 0.5895 | FID: 135.296586 | FID Runtime: 8.10 sec

🌈 Train

Step1: Prepare training data

We use LSDIR (84991) and the first 10k images from FFHQ as our training data. Please write their absolute paths into a txt file. Each line in the txt file should be the absolute path of an image. Just like:

LSDIR/HR_merge/0000001.png
LSDIR/HR_merge/0000002.png
...

Step2: Training for OSEDiff

Pleasr put your txt file path at YOUR TXT FILE PATH. If you have 4 GPUs, you can run

CUDA_VISIBLE_DEVICES="0,1,2,3," accelerate launch train_osediff.py \
    --pretrained_model_name_or_path=SD21BASE_PATH \
    --ram_path=RAM_PATH \
    --learning_rate=5e-5 \
    --train_batch_size=4 \
    --gradient_accumulation_steps=1 \
    --enable_xformers_memory_efficient_attention --checkpointing_steps 500 \
    --mixed_precision='fp16' \
    --report_to "tensorboard" \
    --seed 123 \
    --output_dir=experience/osediff \
    --dataset_txt_paths_list 'YOUR TXT FILE PATH','YOUR TXT FILE PATH' \
    --dataset_prob_paths_list 1,1 \
    --neg_prompt="painting, oil painting, illustration, drawing, art, sketch, cartoon, CG Style, 3D render, unreal engine, blurring, dirty, messy, worst quality, low quality, frames, watermark, signature, jpeg artifacts, deformed, lowres, over-smooth" \
    --cfg_vsd=7.5 \
    --lora_rank=4 \
    --lambda_lpips=2 \
    --lambda_l2=1 \
    --lambda_vsd=1 \
    --lambda_vsd_lora=1 \
    --deg_file_path="params_realesrgan.yml" \
    --tracker_project_name "train_osediff"

📷 Results

• For convenient evaluation and comparison, we have published the test results of DIV2K_val, RealSR, and DRealSR on Google Drive.

Quantitative Comparisons (click to expand)

Visual Comparisons (click to expand)

🎫 License

This project is released under the Apache 2.0 license.

📧 Contact

If you have any questions, please feel free to contact: [email protected]

🎓Citations

@article{wu2024one,
  title={One-Step Effective Diffusion Network for Real-World Image Super-Resolution},
  author={Wu, Rongyuan and Sun, Lingchen and Ma, Zhiyuan and Zhang, Lei},
  journal={arXiv preprint arXiv:2406.08177},
  year={2024}
}

statistics