Reproduce “Data-free Distillation with Degradation-prompt Diffusion for Multi-weather 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
Pei Wang, Xiaotong Luo, Yuan Xie, Yanyun Qu
Researchers from Xiamen University introduce D4IR, a data-free knowledge distillation framework that leverages degradation-prompt diffusion models to synthesize domain-related degraded images for multi-weather image restoration, achieving efficient lightweight student models with performance comparable to data-present methods.
Key insights
- Pre-trained conditional diffusion models can effectively synthesize high-quality, domain-related degraded images for data-free knowledge distillation in low-level vision, overcoming limitations of GAN-based approaches.
- Disentangling degradation information (via learned prompts) from content information (via clean image latents) within the diffusion process is effective in mitigating domain shift and enhancing the relevance of generated pseudo-data.
- Jointly optimizing the degradation prompt adapter and the student network ensures that the generated degradation prompts are not only diverse but also directly contribute to effective knowledge transfer from the teacher to the student.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.