UHD Image Dehazing via anDehazeFormer with Atmospheric-aware KV Cache

TL;DR

anDehazeFormer tackles UHD image dehazing by fusing physics-informed priors with a transformer backbone. The method introduces an Atmospheric Parameter Estimator and a physics-guided KV caching mechanism with adaptive normalization to accelerate training and reduce memory, while a PA-STB and multi-scale reconstruction preserve details under haze constraints. A Physics-Aware Attribution Map (PAAM) integrates physical and feature gradients to improve interpretability. Empirical results on 8K/4K datasets show state-of-the-art restoration with real-time UHD throughput and favorable efficiency metrics, suggesting practical applicability for large-scale dehazing tasks.

Abstract

In this paper, we propose an efficient visual transformer framework for ultra-high-definition (UHD) image dehazing that addresses the key challenges of slow training speed and high memory consumption for existing methods. Our approach introduces two key innovations: 1) an \textbf{a}daptive \textbf{n}ormalization mechanism inspired by the nGPT architecture that enables ultra-fast and stable training with a network with a restricted range of parameter expressions; and 2) we devise an atmospheric scattering-aware KV caching mechanism that dynamically optimizes feature preservation based on the physical haze formation model. The proposed architecture improves the training convergence speed by \textbf{5 $\times$} while reducing memory overhead, enabling real-time processing of 50 high-resolution images per second on an RTX4090 GPU. Experimental results show that our approach maintains state-of-the-art dehazing quality while significantly improving computational efficiency for 4K/8K image restoration tasks. Furthermore, we provide a new dehazing image interpretable method with the help of an integrated gradient attribution map. Our code can be found here: https://anonymous.4open.science/r/anDehazeFormer-632E/README.md.

Figures (7)

• Figure 1: Improvement of our model over the SOTA approaches in 8KDehaze chen2025tokenize0. The circle size is proportional to the number of model parameters.
• Figure 2: The framework integrates four key components. (a) Atmospheric Estimator generating transmission maps and global light parameters, (b) KV Cache Update Module dynamically retaining features via atmospheric-guided retention policies, (c) Parameter-Aware Swin Blocks fusing physical constraints during multi-scale reconstruction, (d) Physics-Aware Attribution Map (PAAM) quantifying decision rationales through gradient integration. Modules are synergistically linked, Atmospheric parameters guide cache retention in (b) and feature restoration in (c), while PAAM (d) traces gradient flows from (a)-(c) to enhance interpretability.
• Figure 3: Illustration of the baseline image and the path function.
• Figure 4: Dehazed results on the O-HAZE Oancuti2018haze and I-HAZE Iancuti2018haze dataset. The proposed anDehazeFormer demonstrates higher color fidelity and restores more details compared with other SOTA methods.
• Figure 5: Dehazed results on the 4KID zheng2021ultra dataset. The proposed anDehazeFormer effectively avoids ghosting and blurring, achieving full-resolution inference with remarkable clarity and precision.