• [2026/03/26] We have released a new benchmark for real-world image degradation, RealIR-Bench.
• [2026/03/26] We have released the RealRestorer data pipeline and model weights.

• Open-source the inference code, degradation pipeline code, and RealIR-Bench evaluation code.
• Release RealIR-Bench on Hugging Face.
• Open-source the RealRestorer model weights.
• Open-source the Qwen-Image-Edit-2511 version.

This project relies on the patched local diffusers/ checkout in this repository.

• Python: 3.12

python3.12 -m pip install --upgrade pip

cd diffusers
python -m pip install -e .

cd ..
python -m pip install -r requirements.txt
python -m pip install -e .
python -m pip install -r RealIR-Bench/requirements.txt

You can verify the environment with:

python -c "from diffusers import RealRestorerPipeline; print(RealRestorerPipeline.__name__)"

• Device: cuda
• Torch dtype: bfloat16
• Inference steps: 28
• Guidance scale: 3.0
• Recommended seed: 42

For practical deployment, we recommend using inputs around 1024 x 1024, which offers a good balance between restoration quality, runtime, and memory usage in the current release. Under the recommended configuration above, the peak GPU memory usage is about 34 GB.

python3 infer_realrestorer.py \
  --model_path /path/to/realrestorer_bundle \
  --image /path/to/input.png \
  --prompt "Restore the details and keep the original composition." \
  --output /path/to/output.png \
  --device cuda \
  --torch_dtype bfloat16 \
  --num_inference_steps 28 \
  --guidance_scale 3.0 \
  --seed 42

degradation_pipeline/ is the synthetic degradation pipeline used by RealRestorer. It is released together with this repository and can be used to synthesize real-world degradations for training, analysis, and controlled evaluation.

The current pipeline covers common restoration settings including blur, haze, noise, rain, moire, and reflection.

Generation script:

python3 infer_degradation.py \
  --image /path/to/input.png \
  --degradation reflection \
  --output /path/to/degraded.png \
  --reflection_ckpt_path /path/to/130_net_G.pth

The script writes the degraded image and a JSON metadata file for the sampled degradation settings.

RealIR-Bench uses a VLM-based scoring protocol to compare restored images against references and reports LPIPS_Score, VLM_Score_Diff, and FS, where:

FS = 0.2 * VLM_Score_Diff * (1 - LPIPS)

Evaluation script:

python3 evaluate_realir_bench.py \
  --ref-dir /path/to/reference_dir \
  --pred-dir /path/to/prediction_dir \
  --task reflection \
  --vlm-model-path /path/to/Qwen3-VL-8B-Instruct

For degradation-specific evaluation, set --task to the corresponding restoration target such as blur, noise, rain, moire, or reflection.

If you find RealRestorer useful in your research, please star and cite:

@misc{yang2026realrestorergeneralizablerealworldimage,
      title={RealRestorer: Towards Generalizable Real-World Image Restoration with Large-Scale Image Editing Models},
      author={Yufeng Yang and Xianfang Zeng and Zhangqi Jiang and Fukun Yin and Jianzhuang Liu and Wei Cheng and jinghong lan and Shiyu Liu and Yuqi Peng and Gang YU and Shifeng Chen},
      year={2026},
      eprint={2603.25502},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2603.25502},
}