P-PatchDiff: Progressive Patch Diffusion Models for Low-light Image Enhancement

TL;DR

P-PatchDiff introduces a progressive patch diffusion framework that dynamically increases patch size during denoising to balance local detail and global brightness context. It uses a multi-patch alignment strategy with a global brightness proxy to maintain visual coherence across varying illumination levels. The method achieves state-of-the-art enhancement across diverse image sizes while being 80 times faster and using under 9GB of memory.

Why It Matters

Low-light image enhancement is critical for applications like surveillance, mobile photography, and medical imaging, where image quality directly impacts decision-making. P-PatchDiff enables efficient, scalable enhancement without sacrificing visual consistency, making it practical for real-world deployment on devices with limited resources.

Key Findings

• Progressive patch size increase allows the model to transition from local to global context during denoising, improving brightness and color estimation.
• Multi-patch alignment with a global brightness proxy ensures consistent illumination across patches of different scales.
• The model achieves high-quality enhancement across image resolutions from 400×600 to 4K with 80× speedup and under 9GB memory usage.