⚡️RAR (Restore, Assess, Repeat): A Unified Framework for Iterative Image Restoration

Official repository for Restore, Assess, Repeat: A Unified Framework for Iterative Image Restoration

Project Page | Paper | Video | Code

Updates

• April 2026: ✨ Source code has been released!
• February 2026: ✨ RAR was accepted into CVPR 2026!

RAR

RAR integrates and within a shared latent space, enabling an iterative Restore–Assess–Repeat process for high-quality image restoration.

• LQA performs fine-grained quality assessment and degradation analysis directly in the latent space, providing adaptive feedback for restoration.
• UIR leverages diffusion priors to perform unified image restoration, handling diverse and composite degradations in a single model.

By tightly coupling assessment and restoration, RAR dynamically identifies degradations and refines outputs, achieving strong performance under unknown and composite degradations.

Setup

1. Clone and enter into repo directory.

git clone https://github.com/SamsungLabs/RAR.git
cd RAR

2. Create a conda environment and activate it.

./environment_setup.sh rar

3. Download pretrained UniRestore checkpoints and place them into path (./checkpoints/).

• Quickly download all checkpoints from Model_Zoo, or download individual checkpoints as needed.

cd checkpoints/
python download.py

• (option) Please update the configuration as follows:

  • configs/infer_cfg.yaml:

    • data.data_dir
    • model.model_pretrained
    • model.load_from
    • vae.vae_pretrained
    • text_encoder.text_encoder_pretrained
    • connector.load_from
    • sentence_model

  • run.py

    • RARInference.model_path
    • RARInference.sentence_model

  • iqa/config.yaml

    • model.vision_encoder_path
    • model.tgt_vision_projector_weight
    • model.llm_path
    • model.delta_path
    • model.checkpoints

  • configs/sd35_imgflow_config/ir_sd35m_r512_multi.yaml (for training)

    • data.data_dir
    • model.model_pretrained
    • model.load_from
    • vae.vae_pretrained
    • text_encoder.text_encoder_pretrained
    • connector.load_from

Run Demo (fast)

CUDA_VISIBLE_DEVICES=0 streamlit run run.py

📌 Open the provided URL in your browser and wait for the model to finish loading.
📌 Click 'Upload' to use your own image or 'Select' a sample to begin the RAR process.
📌 Move the 'progress bar' to compare intermediate outputs.
💡 Feel free to use the samples provided in the /demo_sample directory.

Prepare Dataset

We collect datasets for single, unknown, and composite degradations. Each dataset is organized in JSON format. Please download the datasets into the corresponding folders, and then use the following command to generate the data list:

cd dataset
python download.py
bash decomp_all.sh
python scripts/0_generate_list.py
python scripts/1_generate_iqa_brief.py

Data Structures:

DATA_ROOT/PIR_tar/
    ├── $task_name/
    │   ├── $dataset/
    │   │   ├── image_folder/
    │   │   ├── metas/

DATA_ROOT is the path containing the counting datasets; Each dataset/metas should include: a subset.list specifying the image list and a subset.json containing the formatted metadata.

Inference

• Find more details in ./scripts, change "src_dir" in ./scripts/test_autocomparison.sh

bash scripts/test_autocomparison.sh 0,1,2,3,4,5,6,7 configs/infer_cfg.yaml checkpoints/RAR_modelzoo/RAR/checkpoints/latest.pth output_composite/results/ 4 4 256 online d2c multi_c2c_d2c_online_compare_s4r4 bf16 

Train

For faster reproduction, we recommend using the pre-trained LQA model and training the framework end-to-end.

bash train_scripts_imgflow/train_imgflow_e2e_multi.sh \
8 \
output_imgflow/sd35m_multi_single_bf16 \
configs/sd35_imgflow_config/ir_sd35m_r512_multi.yaml \
--data.data_dir="[dataset/PIR_tar]" \
--data.type=IRImgDataset \
--model.multi_scale=false \
--train.train_batch_size=1 \
--train.gradient_accumulation_steps=16 \
--data.image_size=256 \
--model.image_size=256 \
--data.dset=train_brief \
--data.max_samples=null \
--model.load_from="./checkpoints/RAR_modelzoo/backbone/epoch_100_step_302601.pth" 

Acknowledgements

This work builds upon the following open-source projects:

• Sana: Efficient High-Resolution Image Synthesis with Linear Diffusion Transformer, In ICLR 2025 Oral. [GitHub, Paper]
• Stable Diffusion 3.5: Scaling Rectified Flow Transformers for High-Resolution Image Synthesis, In ICML 2024. [GitHub, Paper]
• DepictQA: Depicted Image Quality Assessment with Vision Language Models, In ECCV 2024. [GitHub, Paper]

We sincerely thank the authors for their open-source contributions.

📖BibTeX

@article{chen2026restore,
  title={Restore, Assess, Repeat: A Unified Framework for Iterative Image Restoration},
  author={Chen, I and Hadji, Isma and Sanchez, Enrique and Bulat, Adrian and Kuo, Sy-Yen and Timofte, Radu and Tzimiropoulos, Georgios and Martinez, Brais and others},
  journal={arXiv preprint arXiv:2603.26385},
  year={2026}
}