Reproduce “AGLLDiff: Guiding Diffusion Models Towards Unsupervised Training-free Real-world Low-light Image 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
Yunlong Lin, Tian Ye, Sixiang Chen, Zhenqi Fu, Yingying Wang, Wenhao Chai, Zhaohu Xing, Lei Zhu and 1 more
AGLLDiff introduces an unsupervised and training-free framework for real-world low-light image enhancement, guiding a pre-trained unconditional diffusion model using a novel attribute-guidance strategy. The method achieves state-of-the-art quantitative performance and visually superior results across diverse synthetic and real-world datasets by focusing on desired image characteristics like exposure, structure, and color.
Key insights
- Shifting the problem paradigm from modeling unknown degradations to guiding towards easily accessible desired attributes of high-quality images enables truly unsupervised and training-free enhancement.
- Attribute guidance, comprising exposure, structure, and color fidelity constraints, provides sufficient information to steer an unconditional diffusion model to robustly enhance real-world low-light images.
- Dynamically adjusting the guidance strength and steps during the diffusion process significantly improves the controllability and effectiveness of classifier guidance, leading to higher fidelity and reduced artifacts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.