Classic Video Denoising in a Machine Learning World:
Robust, Fast, and Controllable

Xin Jin^1,* Simon Niklaus^2,† Zhoutong Zhang³ Zhihao Xia² Chunle Guo^1,# Yuting Yang³ Jiawen Chen³ Chong-Yi Li¹

¹VCIP, CS, Nankai University    ²Adobe Research    ³Adobe
^*This work was done when Xin Jin was an intern at Adobe Research
^†Project Lead    ^#Corresponding Author

CVPR 2025

Paper (TBD) Code (TBD) BibTex

🔥 General Video Denoising with Real-Time Performance! 🔥

Nearly 32 FPS on a single RTX 3090 GPU with HD video inputs!

Compared to previous SOTA video denoising methods (all methods are trained with same noise model), our method could effectively remove out-of-model noise from various videos with real-time performance!

† means we retrain those models with our proposed degradation pipeline.
'teaser_video' is provided by Robert Kjettrup

🤖️ Controllable Ability to Remove or Perserve Noise

With classical video denoising methods, we can control the denoising strength for videos.
Best viewed with maximum contrast and minimum saturation.

This footage is shoted for movie 'Spiral' (2013)

BibTex


@inproceedings{jin2025classic,
    title={Classic Video Denoising in a Machine Learning World: Robust, Fast, and Controllable},
    author={Jin, Xin and Niklaus, Simon and Zhang, Zhoutong and Xia, Zhihao and Guo, Chunle and Yang, Yuting and Chen, Jiawen and Li, Chong-Yi},
    journal={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
    year={2025}
}

Contact

Feel free to contact us at xjin[AT]mail.nankai.edu.cn!

Visitor Count