Reproduce “Bird-SR: Bidirectional Reward-Guided Diffusion for Real-World Image Super-Resolution” 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
Zihao Fan, Xin Lu, Yidi Liu, Jie Huang, Dong Li, Xueyang Fu, Baocai Yin
Bird-SR introduces a bidirectional reward-guided diffusion framework for real-world image super-resolution, jointly leveraging synthetic paired data and unpaired real-world low-resolution images. The framework generates perceptually realistic and structurally consistent high-resolution images, achieving superior perceptual quality metrics and user preferences over existing methods on real-world benchmarks while preserving structural fidelity and suppressing artifacts.
Key insights
- Jointly leveraging both synthetic LR-HR pairs and real-world unpaired LR images through a bidirectional optimization framework effectively bridges the distribution gap and enhances generalization for real-world super-resolution.
- Mitigating reward hacking in perception-driven diffusion models can be achieved through relative reward formulations for paired data and semantic alignment constraints for unpaired real-world data, ensuring both perceptual gains and structural integrity.
- Dynamically adjusting the balance between distortion (fidelity) and perceptual quality across diffusion timesteps is crucial for optimal results, as different steps contribute distinctly to structure and texture synthesis.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.