Reproduce “Test-Time Degradation Adaptation for Open-Set 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
Yuanbiao Gou, Haiyu Zhao, Boyun Li, Xinyan Xiao, Xi Peng
TAO restores images with degradation types absent from pretraining by adapting a small degradation adapter to each damaged test image while a fixed diffusion model generates and guides plausible clean images. On 20 HSTS dehazing images, TAO reports 22.29 PSNR and 0.877 SSIM, exceeding zero-shot YOLY but below supervised DehazeNet; settings are tuned by degradation type.
Key insights
- Open-Set Image Restoration can be addressed by treating it as a problem of unidentified distribution shifts between pretraining and testing, for which Test-Time Adaptation is a suitable solution.
- Pre-trained image diffusion models offer a robust, degradation-agnostic foundation due to their rich generative knowledge, capable of producing high-quality clean images.
- Effectively guiding the diffusion process requires a test-time adaptable degradation model and a multi-stage guidance strategy aligned with the diffusion model's temporal dynamics.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.