[ICLR-2026] RestoreVAR

Sudarshan Rajagopalan | Kartik Narayan | Vishal M. Patel

Code for RestoreVAR: Visual Autoregressive Generation for All-in-One Image Restoration.

Getting Started

Environment

Create the environment as follows

conda create -n var_test python=3.9 -y
conda activate var_test
pip install -r requirements.txt

Downloads

Download checkpoints file and extract it to ckpts/.

Download training and testing data and extract them to data/ and test_data/, respectively.

Download meta-info json files for testing and extract them to test_jsons/.

Testing

To get results of Table 1, run

bash test_regular.sh

To get results of Table 2, run

bash test_generalization.sh
bash metric_generalization.sh

The scripts can be modified as needed.

Training

For training the VAR transformer for restoration, run

bash train.sh

The script loads the pretrained ckpts/var_d16.pth model and introduces the proposed components to train the model for restoration.

Once trained, the latent refiner transformer (LRT) can be trained using

bash train_refiner.sh

Prior to running this command, rename the existing local_output/ directory to something else. The refiner uses the ckpts/vae_restorevar.ckpt file which contains weights for the VAE decoder fine-tuned to handle continuous latents.

The above commands use trainset.json and valset.json which contain information about file paths, datasets, etc. To include your own datasets, make .json files with entries as follows:

{
    "image_path": <degraded_image_path>, 
    "target_path": <target_image_path>, 
    "degradation": <degradation>, 
    "dataset": <dataset>
}

Citation

If you found our work useful, please cite:

@inproceedings{
    rajagopalan2026restorevar,
    title={Restore{VAR}: Visual Autoregressive Generation for All-in-One Image Restoration},
    author={Sudarshan Rajagopalan and Kartik Narayan and Vishal M. Patel},
    booktitle={The Fourteenth International Conference on Learning Representations},
    year={2026},
    url={https://openreview.net/forum?id=yvXtCn2zfz}
}

Acknowledgments

Our code uses parts from VAR and VARSR. We thank the authors for sharing their codes!