Reproduce “InstructIR: High-Quality Image Restoration Following Human Instructions” 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
Marcos V. Conde, Gregor Geigle, Radu Timofte
InstructIR introduces the first all-in-one image restoration model capable of high-quality image recovery guided by natural human instructions. The system achieves an average PSNR of 29.55 dB on a 5-degradation benchmark, representing a 1dB improvement over previous state-of-the-art all-in-one methods, and enables controllable enhancement through sequential text prompts.
Key insights
- Natural language instructions provide a more intuitive and flexible guidance mechanism for all-in-one image restoration compared to abstract learned prompts or explicit degradation labels.
- A compact, pure text-based sentence encoder is sufficient for effectively capturing the semantic intent of restoration instructions, contributing to computational efficiency.
- Dynamic, instruction-conditioned modulation of image features allows a single unified model to effectively handle multiple and diverse image restoration tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.