Reproduce “Generative Diffusion Prior for Unified Image Restoration and Enhancement” 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
Ben Fei, Zhaoyang Lyu, Liang Pan, Junzhe Zhang, Weidong Yang, Tianyue Luo, Bo Zhang, Bo Dai
Generative Diffusion Prior restores images from known or unknown degradations by reusing one unconditional ImageNet-trained diffusion model. At each denoising step, it guides the model’s predicted clean image toward the damaged input and can adjust degradation parameters during sampling. On 25%-inpainting ImageNet tests, its clean-image guidance reached FID 16.58 and consistency 5.29, but DDRM had better consistency.
Key insights
- Guidance is more effective when computed in the clean-image domain: GDP-x₀ consistently outperforms guidance on xₜ, supporting the need to match the noise characteristics of the observation and the diffusion estimate.
- A fixed unconditional diffusion prior can support substantially different restoration problems by changing only the degradation model and inference-time objectives, including blind and nonlinear cases.
- The framework favors perceptual plausibility and observation consistency over pixel-perfect reconstruction, so improved FID or LPIPS may accompany lower PSNR or SSIM because generated details can differ from the reference.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.