Semantic-guided diffusion for water-related image enhancement

Open Webpage

Jingchun Zhou, Jiatong Liu, Dehuan Zhang, Xiuguo Zhang, Zifan Lin

Published: 2026, Last Modified: 09 Nov 2025Pattern Recognit. 2026EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: Highlights•Pseudo-label guided diffusion enables label-free underwater enhancement.•U-Net diffusion with linear scheduling restores global and local details.•Semantic prompts fused with Zip-CLIP heatmaps guide pseudo-supervision.•G-CNN integrates cues to improve sharpness and semantic structure.•Outperforms state-of-the-art on UIEB, U45, and UCCS benchmarks.