PAWS

🐾 PAWS: Practical Video Super-Resolution

PAWS is a modular video super-resolution framework for training, evaluating, publishing, and running practical VSR models. This model repository hosts PAWS RealBasicVSR++ weights for x4 real-world video super-resolution, including the recommended HAT-LPIPS-TRes model and supporting ablation checkpoints.

For setup, inference commands, configuration details, and app usage, see the PAWS code repository:

https://github.com/jnguyen5650/PAWS

📦 Model Files

EMA files are raw generator weights for test.py and research workflows. The .paws.pth file is a portable artifact created with tools.publish_model for the PAWS demo app. Only the recommended LPIPS-TRes model is released in the app artifact format.

🚧 Limitations

• The released models target x4 super-resolution.
• Results depend on the degradation patterns present in the input video.
• The models may hallucinate high-frequency texture under severe degradation.
• Large inputs may require tiled inference to fit available GPU memory.

📝 Citation

If you use these weights or the PAWS codebase, please cite the PAWS preprint:

@article{nguyen2026paws,
  title={PAWS: Practical Real-World Video Super-Resolution with Modular Cleaning},
  author={Nguyen, Justin},
  journal={Preprint},
  year={2026}
}

⚖️ License and Acknowledgements

The PAWS code and released model materials are provided under the Apache 2.0 license. See the PAWS repository NOTICE.md file for acknowledgements and third-party code attributions.