Hihaze speeds up and improves high-resolution image dehazing
Hihaze: One-Step Diffusion for High-Resolution Real-World Image Dehazing
Computer Vision and Pattern Recognition
Summary
Haze makes photos look foggy and unclear, especially in big pictures with uneven haze in different spots. The authors present Hihaze, a new method that clears up high-resolution images in one step by carefully using information about where haze is and how strong it is. They also designed a way to combine smaller parts of the image smoothly without extra training. Tests show Hihaze makes clearer pictures faster than previous methods.
What this means in practice
- •For photography editors: Restore clear outdoor images taken in hazy conditions without manual adjustments, maintaining structural details in high-resolution photos.
- •For security monitoring teams: Improve visibility in high-resolution video or image feeds affected by varying haze, enabling better analysis of real-world scenes.
Authors
Zheng Chen, Yuchen Huang, Xiaoyang Liu, Jue Gong, Linghe Kong, Yulun Zhang
Abstract
Diffusion models have recently achieved encouraging results in real-world image dehazing. However, restoring high-resolution images remains challenging due to spatially non-uniform haze and inconsistent predictions across local patches. To address these issues, we propose Hihaze, a one-step diffusion framework for high-resolution real-world dehazing. Hihaze adopts two haze-aware modulation modules: attentive skip compensation (ASC) and a spatial degradation adapter (SDA). ASC selectively transfers reliable encoder information to the variational autoencoder (VAE) decoder, preserving structural details while suppressing haze-degraded features. SDA extracts multi-scale spatial degradation representations and injects them into the U-Net to guide restoration according to local haze variations. Furthermore, we design a training-free global-guided patch fusion (GPF) method. It uses a global prediction to align patch appearance and downweight unreliable predictions, followed by structure-guided detail compensation. Notably, this training-free design also makes GPF a plug-and-play module for other dehazing models. Experiments across four real-world benchmarks show that Hihaze surpasses state-of-the-art approaches in both visual quality and inference speed. Specifically, it delivers PSNR improvements of 3.62~dB on Dense-Haze and 3.69~dB on NH-HAZE compared with the strongest baselines. Code: https://github.com/zhengchen1999/Hihaze.