Unveiling Silent Hill with HeteroDiff: Dense Haze Removal via Heterogeneous Diffusion

Abstract Dense haze removal remains highly challenging due to severe information loss and heterogeneous degradation. Although recent diffusion‐based methods have shown promising performance in image dehazing, they usually overlook the heterogeneous restoration difficulty induced by dense haze degradation, resulting in imbalanced optimization and unstable generation. To address these issues, we propose HeteroDiff , a novel heterogeneous diffusion framework for dense haze removal. Specifically, we exploit the heterogeneous degradation characteristics of dense haze images and devise an Interleaved Curriculum Learning (ICL) strategy that jointly schedules haze densities and diffusion timesteps, enabling the model to progressively learn restoration tasks from easy to hard for balanced optimization and accelerated convergence. Moreover, we incorporate depth priors into a Depth‐Enhanced Noise Predictor (DENP) to improve structural restoration under severe haze degradation. During reverse diffusion sampling, a Context‐Aware Sampling (CAS) strategy is further introduced to suppress hallucinated artifacts and improve spatial consistency by leveraging broader contextual information. Extensive experiments on NH‐HAZE and DENSE‐HAZE demonstrate that HeteroDiff consistently outperforms existing state‐of‐the‐art methods in both visual quality and downstream object detection performance, particularly under severe haze conditions.

Authors

Institutions

Publication Details

Journal
Computer Graphics Forum
Published
2026-11-13
DOI
https://doi.org/10.1111/cgf.70616
Primary Topic
Image Enhancement Techniques
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Unveiling Silent Hill with HeteroDiff: Dense Haze Removal via Heterogeneous Diffusion

Xiao Lv, Ying Yang, Tao Xiang, Bin Zhu et al.
5 links
Computer Graphics Forum
Image Enhancement Techniques
article

Unveiling Silent Hill with HeteroDiff: Dense Haze Removal via Heterogeneous Diffusion

Xiao Lv, Ying Yang, Tao Xiang, Bin Zhu, Chuan Ma, Shengfeng He
article en

Abstract

Abstract Dense haze removal remains highly challenging due to severe information loss and heterogeneous degradation. Although recent diffusion‐based methods have shown promising performance in image dehazing, they usually overlook the heterogeneous restoration difficulty induced by dense haze degradation, resulting in imbalanced optimization and unstable generation. To address these issues, we propose HeteroDiff , a novel heterogeneous diffusion framework for dense haze removal. Specifically, we exploit the heterogeneous degradation characteristics of dense haze images and devise an Interleaved Curriculum Learning (ICL) strategy that jointly schedules haze densities and diffusion timesteps, enabling the model to progressively learn restoration tasks from easy to hard for balanced optimization and accelerated convergence. Moreover, we incorporate depth priors into a Depth‐Enhanced Noise Predictor (DENP) to improve structural restoration under severe haze degradation. During reverse diffusion sampling, a Context‐Aware Sampling (CAS) strategy is further introduced to suppress hallucinated artifacts and improve spatial consistency by leveraging broader contextual information. Extensive experiments on NH‐HAZE and DENSE‐HAZE demonstrate that HeteroDiff consistently outperforms existing state‐of‐the‐art methods in both visual quality and downstream object detection performance, particularly under severe haze conditions.

Computer Graphics Forum
Chongqing University (CN), Singapore Management University (SG)
National Natural Science Foundation of China
Image Enhancement Techniques
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon