Clear Nights Ahead: Towards Multi-Weather Nighttime Image Restoration

Conference: AAAI 2026 arXiv: 2505.16479 Code: https://henlyta.github.io/ClearNight/ Area: Image Restoration / Nighttime Adverse Weather Removal Keywords: nighttime image restoration, multi-weather, Retinex prior, dynamic MoE, AllWeatherNight dataset

TL;DR

This paper is the first to define and explore the multi-weather nighttime image restoration task. It constructs the AllWeatherNight dataset (8K training + 1K synthetic test + 1K real-world test) and proposes the ClearNight unified framework, which simultaneously removes compound degradations—haze, rain streaks, raindrops, snow, and flare—in a single stage via Retinex dual-prior guidance and weather-aware dynamic specificity–commonality collaboration. With only 2.84M parameters, ClearNight comprehensively surpasses state-of-the-art methods.

Background & Motivation

In nighttime scenes, adverse weather degradations are tightly coupled with non-uniform illumination (flare effects, halos), severely impacting downstream tasks such as autonomous driving and video surveillance. Three major gaps exist in prior research:

Gap	Specific Issue	Representative Works
Dataset absence	No dataset covering multi-weather + nighttime + flare simultaneously	UNREAL-NH (haze only), GTAV-NightRain (rain only)
Method limitation	Nighttime methods handle only a single degradation type	TKL (dehazing), FSDGN (deraining)
Neglected degradation coupling	Illumination and weather degradations are intertwined (e.g., rain streaks appear brighter under lights, haze is denser in dark regions)	Daytime methods WeatherDiff, AWRaCLe cannot handle this

ClearNight must address two core challenges: (1) the lack of realistic multi-weather nighttime training samples; and (2) the inability of existing models to effectively handle coupled degradations. A cascade approach (dehazing followed by deraining) yields only 27.8 dB PSNR and is slow.

Method

Overall Architecture

ClearNight adopts a DehazeFormer-based encoder–decoder architecture integrating two core modules: Retinex dual-prior guidance (decoupling illumination from texture) and weather-aware dynamic specificity–commonality collaboration (WDS + TFB dual-branch processing of multi-weather degradations). The entire model contains only 2.84M parameters.

Key Designs

1. AllWeatherNight Dataset Construction

   • 2,000 high-quality nighttime images are selected from BDD100K and ExDark as ground truth (triple filtering by brightness/gradient/variance + manual secondary screening).
   • Illumination-aware degradation synthesis: Flare is first simulated by convolving with the Atmospheric Point Spread Function (APSF) (\(X^\text{flare} = \alpha X + \beta(L * K^\text{APSF})\), where \(\beta\) is adaptively set according to the light source pixel ratio), followed by illumination-aware weather degradation synthesis (\(X^d = X^\text{flare} + \sum_{e \in \mathcal{E}} \omega_e \cdot \mathcal{F}_e^G(X^\text{flare})\)).
   • Key innovation: The weight map \(\omega_e\) uses the illumination map from Retinex decomposition, naturally coupling weather degradation intensity with illumination (except raindrops, where \(\omega_\text{RD}=1\), as they are governed by local background).
   • t-SNE validation: The illumination-aware synthetic data distribution is substantially closer to real-world data than conventional uniform synthesis.

2. Retinex Dual-Prior Guidance

   • The degraded image is decomposed via Retinex into a reflectance component \(R^d\) and an illumination component \(I^d\) (\(X^d = R^d \cdot I^d\)).
   • Illumination prior \(i_n\): Injected into the first three Transformer Feature Blocks (TFBs) to guide the network toward non-uniform illumination regions.
   • Reflectance prior \(r_n\): Injected into WDS blocks to enhance texture representation, aiding in distinguishing weather degradation types and recovering background details.
   • A shared-weight multi-scale prior extraction unit (MPE) extracts three-scale priors via dilated convolutions.

3. Weather-Aware Dynamic Specificity–Commonality Collaboration

   • Commonality branch: Sequential Transformer Feature Blocks (TFBs) capturing shared features across all weather degradations.
   • Specificity branch (WDS): Dynamically selects the top-10 units from 25 candidates to adaptively construct a sub-network.
   • Weather Guider: Performs multi-label classification (BCE loss), learns weather-specific prototypes, and automatically associates different weather types with distinct candidate unit combinations.

Loss & Training

\[L_\text{total} = L_1 + 0.1 \cdot L_\text{perceptual} + 0.001 \cdot L_\text{bce} + 0.01 \cdot L_\text{lb} + 0.02 \cdot L_\text{depth}\]

Adam optimizer, lr=2×10⁻⁴, cosine annealing, 100 epochs, patch size 256×256, 2.84M parameters, 0.32s inference time.

Key Experimental Results

Main Results: AllWeatherNight Synthetic Test Set (Scene-Level)

Scene	Method	PSNR↑	SSIM↑
Rain Scene	TKL	29.09	0.8769
	AWRaCLe	31.54	0.9210
	ClearNight	32.59	0.9223
Snow Scene	RAMiT	29.12	0.8889
	AWRaCLe	29.43	0.8738
	ClearNight	30.65	0.9041
Haze	RAMiT	36.44	0.9738
	ClearNight	36.47	0.9621
Rain Streaks Only	DEA-Net	32.76	0.9285
	ClearNight	33.62	0.9331

Ablation Study

Configuration	PSNR↑	SSIM↑	Note
Baseline (DehazeFormer)	28.80	0.8825	No prior, no dynamic branch
+ Illumination prior	32.13	0.9148	+3.33 dB, largest single contribution
+ Reflectance prior	32.39	0.9189	Texture enhancement effective
+ Dynamic routing	32.49	0.9207	WDS dynamic routing effective
+ Weather Guider	32.59	0.9223	Full ClearNight

Key Findings

• The illumination prior contributes the most (+3.33 dB), validating the central importance of illumination decoupling in nighttime restoration.
• Real-world evaluation: rain streak scene NIQE 3.7653 (best), snow scene NIQE 3.2191 (best).
• Models trained on AllWeatherNight achieve significantly better NIQE on real-world data than models trained on combinations of existing nighttime datasets.
• ClearNight matches or surpasses dedicated single-weather methods, demonstrating the effectiveness of the unified framework.

Highlights & Insights

• The first multi-weather nighttime image restoration framework and dataset, filling a critical research gap.
• The illumination-aware degradation synthesis strategy cleverly employs the Retinex illumination map as degradation weights, substantially improving synthesis realism.
• The WDS Weather Guider's dynamic unit assignment is interpretable, with different weather types automatically activating distinct candidate unit combinations.
• With only 2.84M parameters, ClearNight outperforms large models such as DEA-Net (15M+) and AWRaCLe across multi-weather nighttime scenarios.

Limitations & Future Work

• Performance is limited under extreme dynamic illumination changes (e.g., rapidly flickering light sources).
• The dataset primarily covers driving/detection scenes; indoor nighttime scenes are underrepresented.
• Standalone flare removal performance is modest (PSNR 38.77 vs. RAMiT 43.01) due to insufficient dedicated flare training data.
• Only four predefined degradation types (haze/rain streaks/raindrops/snow) are supported; rare weather conditions such as dust storms are not covered.

Related Work & Insights

Direction	Representative Methods	Difference from Ours
Daytime multi-weather restoration	WeatherDiff, WGWS, AWRaCLe	Neglect nighttime illumination–weather degradation coupling
Nighttime single-weather restoration	TKL (dehazing), FSDGN (deraining)	Handle only single degradation; cannot address composite scenes
Nighttime datasets	UNREAL-NH, GTAV-NightRain, RVSD	Cover only a single weather type; no flare synthesis
Dynamic networks	MoE, dynamic filtering	ClearNight's WDS integrates weather classification for semantically guided dynamic routing

Rating

• Novelty: ⭐⭐⭐⭐ First to define the multi-weather nighttime restoration task; dataset and illumination-aware synthesis are novel.
• Experimental Thoroughness: ⭐⭐⭐⭐ Evaluation on synthetic + real-world data, ablation study, t-SNE analysis, and cascade comparison.
• Writing Quality: ⭐⭐⭐⭐ Clear paper structure; motivation for illumination-aware degradation synthesis is well articulated.
• Value: ⭐⭐⭐⭐ Fills an important gap; both the dataset and method provide lasting contributions to the community.

Related Papers

• [AAAI 2026] MFmamba: A Multi-function Network for Panchromatic Image Resolution Restoration Based on State-Space Model
• [NeurIPS 2025] MoDEM: A Morton-Order Degradation Estimation Mechanism for Adverse Weather Image Restoration
• [AAAI 2026] Large Language Models Meet Extreme Multi-label Classification: Scaling and Multi-modal Framework
• [AAAI 2026] ClearAIR: A Human-Visual-Perception-Inspired All-in-One Image Restoration
• [ICCV 2025] MP-HSIR: A Multi-Prompt Framework for Universal Hyperspectral Image Restoration