Universal Image Restoration via Internalized Chain-of-Thought Reasoning
Researchers have developed CoTIR, a novel framework for universal image restoration that integrates Chain-of-Thought (CoT) reasoning directly into a single model. This approach aims to overcome the limitations of traditional multi-step CoT methods, such as high computational costs and weak modeling of degradation interactions. By fine-tuning a pre-trained image editing model and encoding CoT-style reasoning through Lagrangian optimization, CoTIR enables holistic restoration without specialized modules. The accompanying CoTIR-Bench, a benchmark with 5.2 million samples, facilitates training and evaluation, demonstrating CoTIR's superior performance in perceptual quality and fidelity compared to existing methods. AI
IMPACT This research introduces a novel approach to image restoration by internalizing complex reasoning, potentially improving efficiency and performance in AI-driven image processing tasks.