Seeing Beyond Haze: Generative Nighttime Image Dehazing

TL;DR

Nighttime images suffer from haze and glow, often obliterating background details. BeyondHaze addresses this by distilling task-specific dehazing priors into a diffusion backbone using LoRA and augmenting training with a detail-enhancement stream and a severe-degradation stream to recover fine-scale details and missing backgrounds. It enables user-controlled generative outputs via prompts and provides a generative-level map to indicate hallucination-prone regions, improving interpretability. On RealNightHaze, BeyondHaze achieves substantial gains in non-reference metrics (e.g., MUSIQ $=65.79$, TRES $=80.08$), demonstrating enhanced visibility and plausible background reconstruction in dense haze.

Abstract

Nighttime image dehazing is particularly challenging when dense haze and intense glow severely degrade or entirely obscure background information. Existing methods often struggle due to insufficient background priors and limited generative capability, both of which are highly important under such conditions. In this paper, we introduce BeyondHaze, a generative nighttime dehazing method that not only reduces haze and glow effects but also reconstructs plausible background structures in regions where visual cues are heavily degraded. Our approach is built on two main ideas: obtaining strong background priors by adapting image diffusion models to nighttime dehazing, and enhancing generative ability in haze- and glow-obscured areas through guided training. Task-specific nighttime dehazing knowledge is distilled into an image diffusion model while preserving its capacity to generate clean images. The diffusion model is further trained on tailored image pairs to improve its ability to recover background details that are suppressed by haze effects. Since generative models may introduce hallucinated content, we design our framework to allow user control over the generative level, enabling a balance between visual realism and fidelity. Experiments on real-world nighttime images demonstrate that BeyondHaze substantially improves visibility and scene detail under dense haze.

Figures (15)

• Figure 1: Qualitative results from NightDeFog'20 yan2020nighttime, NightEnhance'23 jin2023enhancing, SFSNiD'24 cong2024semi and our method on real-world data. Our method not only reduces dense haze and strong glow but also infers missing background details and content in severely degraded regions.
• Figure 2: Our approach integrates dehazing priors with generative capabilities to enhance nighttime image dehazing. We first pretrain a dehazing model using augmentations that simulate noise and light effects in hazy night scenes. The knowledge from this dehazing model is then distilled into a pretrained image diffusion model with LoRA. Additional training pairs are generated by two supplementary models: a detail enhancement model that applies super-resolution to initial dehazed images, and a severe degradation model which produces substantially obscured images from clear images. These training pairs, combined with customized text prompts, are used to fine-tune the diffusion model and enable controllable generative dehazing. During inference, the fine-tuned model performs two inference branches conditioned on low- and high-generative text prompts, producing the corresponding outputs. A generative-level map is then derived by computing attention scores between the two branches, revealing regions that rely more on generative synthesis than on factual restoration.
• Figure 3: Qualitative results from NightEnhance’23 jin2023enhancing, SFSNiD’24 cong2024semi, and our method on real-world datasets. “Ours-low” refers to our low-generative results, which prioritize haze removal, while “Ours-high” refers to our high-generative results, allowing the network to infer details and backgrounds in severely degraded regions. “Gen. Level Map” shows the generative-level maps of “Ours-high,” with red indicating high and blue indicating low generative regions. Zoom in for better visualization.
• Figure 4: Qualitative results from NightDeFog’20 yan2020nighttime, NightEnhance’23 jin2023enhancing, SFSNiD’24 cong2024semi, and our method on real-world datasets. “Ours-low” refers to our low-generative results, which prioritize haze removal, while “Ours-high” refers to our high-generative results, allowing the network to infer details and backgrounds in severely degraded regions. The first two rows exhibit exhibits the ability of “Ours-high” to infer missing scene content, while the last two rows display fine-scale detail generation by Ours-high.
• Figure 5: Qualitative results from NightDeFog'20 yan2020nighttime, NightEnhance'23 jin2023enhancing, SFSNiD'24 cong2024semi and our method on real-world data. Our method not only reduces dense haze and strong glow but also infers missing background details and content in severely degraded regions.