Reproduce “Denoising as Adaptation: Noise-Space Domain Adaptation for Image Restoration” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Kang Liao, Zongsheng Yue, Zhouxia Wang, Chen Change Loy
Researchers from S-Lab, Nanyang Technological University, developed "Denoising as Adaptation," a noise-space domain adaptation framework that leverages a diffusion model during training to enhance the generalization of image restoration networks to real-world degradations without inference overhead. The method achieves state-of-the-art results across denoising, deblurring, and deraining tasks, notably improving denoising performance to 34.71 dB PSNR on the SIDD dataset.
Key insights
- A diffusion model's noise prediction sensitivity to input quality can serve as an effective training signal for a restoration network.
- Adapting images in the "noise space" offers a stable and effective paradigm for domain adaptation, particularly for low-level vision tasks.
- Specific mechanisms, like channel-shuffling and residual-swapping, are essential to prevent proxy models from learning trivial shortcuts when processing mixed-domain inputs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.