ZID-Net improves image dehazing with zero-inference diffusion prior network

Researchers have developed ZID-Net, a novel framework designed to improve single image dehazing by decoupling diffusion supervision from feed-forward inference. This approach aims to combine the strong generative priors of diffusion models with the computational efficiency of CNNs. ZID-Net utilizes a frequency-spatial decoupled backbone with specific blocks for noise filtering and context capture, and a unique Zero-Inference Prior Propagation Head during training to provide supervision without increasing inference cost. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT Introduces a novel architecture for image dehazing that balances quality and efficiency, potentially improving performance in computer vision applications.

RANK_REASON This is a research paper introducing a new technical framework for image processing.

Read on arXiv cs.CV →

• ZID-Net
• RESIDE
• StateHaze1k

• paper
• other