CVPR 2026

From Events to Clarity

The event-guided diffusion framework for dehazing. EvDehaze uses high dynamic range event cues to recover structure, contrast, and temporal detail from hazy dynamic scenes.

Ling Wang^*, Yunfan Lu^*, Wenzong Ma, Huizai Yao, Pengteng Li, Hui Xiong^†
The Hong Kong University of Science and Technology (Guangzhou)

Code Dataset (Coming Soon) CVPR Poster BibTeX

EvDehaze teaser showing hazy RGB, event cues, and restored output

RGB + Event Diffusion Dehazing

First Event-guided diffusion formulation for image dehazing.

HDR Cues Event streams preserve structure where frame contrast is suppressed.

Real Data UAV RGB-event capture platform for real hazy outdoor scenes.

Strong Results Best overall diffusion baseline on SOTS and NH-HAZE metrics.

Abstract

Event cameras reveal the signal that haze hides from frames.

Conventional RGB dehazing struggles when scattering removes scene structures and illumination details. EvDehaze introduces event cameras into dehazing and treats restoration as event-conditioned generation, injecting event-derived edge and illumination features into a latent diffusion model.

Why Events

Events offer high dynamic range sensing and microsecond temporal precision. Under dense haze, they provide sparse but reliable motion, edge, and contrast cues that complement low-contrast RGB frames.

What EvDehaze Adds

The framework aligns hazy RGB frames with temporal event representations, then guides DDIM denoising through cross-attention so the model can restore natural contrast and sharper structures.

Method

RGB restoration steered by event-conditioned diffusion.

EvDehaze keeps the image-generation strength of latent diffusion while adding event guidance at the feature level.

Latent DDIM Backbone A hazy RGB frame is encoded, iteratively refined in latent space, and decoded into the restored image.

Temporal Event Representation Raw events are converted into pyramid features that capture edge, corner, and illumination changes across time.

Cross-Attention Guidance Event features condition intermediate U-Net features through diffusion sampling, improving structure recovery in degraded regions.

Dataset

Real RGB-event haze captures from an airborne platform.

The dataset uses synchronized RGB and event sensors mounted on a UAV to record outdoor scenes under heavy haze. Synthetic RGB-event data is also generated from RESIDE/SOTS for supervised training and evaluation.

Dataset (Coming Soon)

Real-world RGB-event haze dataset collection platform and examples

6,674

Real-world event segments reported in supplementary material.

61,173

Aligned hazy RGB images for real RGB-event evaluation.

Synthetic motion types: radial, rotational, and translational.

Results

Sharper structures and stronger real-world contrast.

Qualitative panels show hazy input, event visualization, EvDehaze output, and histogram comparisons on real UAV captures.

Animated real-world EvDehaze qualitative results

Real-World Visualization

Event cues emphasize scene boundaries and illumination changes that are weak in RGB-only observations, helping the diffusion sampler recover cleaner perceptual structure.

Download MP4 Open Figure

EvDehaze real-world qualitative evaluation

Method	Diffusion	SOTS PSNR	SOTS SSIM	SOTS LPIPS	NH-HAZE PSNR	NH-HAZE SSIM	NH-HAZE LPIPS	Params
IR-SDE	Yes	33.82	0.984	0.014	12.59	0.520	0.361	537.21M
ResShift	Yes	29.06	0.950	0.017	16.26	0.625	0.327	114.65M
EvDehaze	Yes	34.12	0.986	0.012	18.43	0.637	0.313	122.68M

Citation

Reference

Please cite EvDehaze if this project is useful for your research.

@inproceedings{wang2026events,
  title={From Events to Clarity: The Event-Guided Diffusion Framework for Dehazing},
  author={Wang, Ling and Lu, Yunfan and Ma, Wenzong and Yao, Huizai and Li, Pengteng and Xiong, Hui},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={34028--34039},
  month={June},
  year={2026}
}