Source-linked AI summary
CLIP-Guided Data Augmentation for Night-Time Image Dehazing
Xining Ge, Weijun Yuan, Gengjia Chang, Xuyang Li, Shuhong Liu
TL;DR
Nighttime dehazing is difficult because haze, low illumination, and uneven lighting interact under scarce target-domain supervision, while mismatched external data can weaken adaptation. The paper addresses this with CLIP-filtered data curation, two-stage NAFNet training, and inference-time enhancement, reporting stable Y-channel improvements over the hazy-input baseline on NHM-20 while RGB metrics and LPIPS remain below that baseline.
Problem
Nighttime dehazing combines heterogeneous coupled degradations with scarce target-domain samples, and indiscriminate external data may introduce domain mismatch.
Method
The framework uses CLIP similarity to screen auxiliary samples, trains NAFNet first on the target domain and then on filtered broader data, and applies TLC, self-ensemble, and snapshot fusion at inference.
Results
On NHM-20, the framework improves Y-channel PSNR and SSIM over the hazy-input baseline, while RGB metrics and LPIPS do not exceed the baseline.
Takeaways & Limitations
The paper supports a practical pipeline that prioritizes domain-consistent data use, staged optimization, and inference stability over complex backbone redesign.
Abstract
from arXiv · showhide
Nighttime image dehazing faces a more complex degradation pattern than its daytime counterpart, as haze scattering couples with low illumination, non-uniform lighting, and strong light interference. Under limited supervision, this complexity aggravates domain drift and training instability, since target-domain samples are scarce while naively introducing external data may weaken adaptation due to distribution mismatch. This paper presents our solution to the NTIRE 2026 Night Time Image Dehazing Challenge, built as a unified framework that integrates domain-aligned data construction, stage-wise training, and inference-time enhancement. Specifically, a pre-trained CLIP visual encoder screens candidate external samples by similarity to construct training data closer to the target domain. NAFNet is then trained in two stages, first adapting to the target domain and then expanding to broader degradation patterns. At inference time, TLC, x8 self-ensemble, and weighted snapshot fusion are combined to improve output stability. Rather than relying on complex network redesign, the proposed framework offers a practical and effective pipeline for nighttime image dehazing.
1. Introduction
Nighttime image dehazing combines haze scattering with low illumination and uneven artificial lighting, while scarce target data and external-domain mismatch destabilize adaptation. The paper addresses these constraints with CLIP-guided data screening, stage-wise NAFNet training, and inference-time enhancement.
- Motivation: Nighttime degradation couples haze, low illumination, and artificial lighting, causing halo diffusion, dark-region noise, and severe contrast loss.The central challenge is balancing structural recovery, detail preservation, and visual naturalness under heterogeneous lighting.
- Motivation: Limited paired target-domain data and unstable, heterogeneous degradations make simply increasing network capacity insufficient for nighttime dehazing.External data can help, but uncontrolled distribution mismatch may weaken adaptation.
- Approach: The framework screens external samples with a pretrained CLIP visual encoder to retain data similar to the target nighttime domain.This targets both data scarcity and domain shift from external data.
- Approach: NAFNet is trained in two stages: target-domain adaptation first, followed by broader degradation coverage.The staged design is paired with TLC, self-ensemble, and snapshot ensemble during inference to improve stability and reconstruction quality.
- Contribution: The paper presents a unified, task-oriented pipeline rather than relying on complex network redesign.Its co-design spans data screening, optimization, and inference-time enhancement.
2. Related Works
Prior dehazing methods span physical models, CNNs, Transformers, and expressive restoration networks, but their effectiveness often depends on abundant data and stable distributions. Nighttime dehazing instead requires distribution-consistent data use and stable training and inference under coupled, spatially varying degradation.
- Dehazing methods: Physical-prior methods estimate variables such as transmittance, scene irradiance, or air light, while deep models learn hazy-to-clear mappings directly.The literature progresses from interpretable atmospheric-scattering constraints to CNN and Transformer-based restoration.
- Dehazing methods: CNNs, Transformers, and all-in-one restoration models improve detail recovery and global context modeling but usually rely on large datasets and relatively stable distributions.This dependence limits generalization when target data are scarce or degradation distributions change substantially.
- Nighttime dehazing: Nighttime haze couples with low illumination, non-uniform lighting, halos, noise amplification, and color shifts, producing more heterogeneous degradation than daytime haze.Daytime methods consequently transfer poorly to highlight suppression, dark-region detail recovery, and color naturalness.
- Nighttime dehazing: Existing nighttime approaches emphasize detailed imaging models or more expressive networks, yet the bottleneck also involves limited target samples, domain gaps, and training and inference instability.Network expressiveness alone does not address distribution consistency under limited supervision.
- Limited supervision: External data can enlarge training sets but may weaken target-task learning when scene content or degradation patterns mismatch the target domain.Similarity-based screening is presented as a way to expand data while preserving target-domain alignment.
- Robust restoration: Inference-time self-ensemble and model ensemble improve robustness by combining multiple predictions without changing the basic network structure.These strategies address fluctuations caused by individual views or checkpoints in small-sample and high-uncertainty settings.
3. Method
The method builds a target-aligned auxiliary dataset, trains NAFNet in stages, and combines several inference-time enhancements. CLIP screening limits auxiliary samples to those similar to nighttime targets, while staged optimization and ensembling aim to broaden degradation coverage and stabilize outputs.
- Framework overview: The framework proceeds from auxiliary-data screening through two-stage training to test-time enhancement.Its three components are target-aligned data curation, stage-wise restoration training, and inference-time enhancement.
- Cross-dataset data curation: A pretrained CLIP encoder compares external samples with target nighttime images and retains sufficiently similar candidates for the augmented training set.The curation treats auxiliary data as a distribution-aligned extension rather than unconstrained expansion.
- Stage-wise training: Stage one trains NAFNet on the 25 paired NTHazy nighttime samples to establish a target-domain prior.The provided training configuration uses 256×256 patches, batch size 4, and 200K iterations.
- Stage-wise training: Stage two fine-tunes the stage-one model on the combined target and CLIP-filtered data for another 200K iterations.This broadens degradation coverage after the model has acquired nighttime haze and low-light characteristics.
- Inference-time enhancement: Inference combines TLC-style local enhancement, ×8 self-ensemble, and weighted snapshot fusion to improve output robustness.The eight self-ensemble transformations are mapped back and averaged, while snapshots provide complementary checkpoint predictions.
- Inference-time enhancement: Snapshot fusion combines the 80K, 100K, and 200K checkpoints with weights 0.04, 0.01, and 0.95.The final checkpoint dominates, while earlier snapshots act as mild regularizers against checkpoint-specific artifacts.
4. Experiments
The experiments define reproducible training and evaluation settings for nighttime dehazing, then show that the pipeline improves luminance-domain restoration while leaving color and perceptual quality less consistent.
- Qualitative Results: The qualitative comparison examines Input, NAFNet-100k, Weighted Ensemble, and ground truth on representative NHM-20 samples.
- Implementation Settings: The model uses NAFNet/NAFNetLocal with a two-stage training setup and inference-time enhancements summarized in the implementation settings.Stage one adapts on NTHazy, while stage two adds CLIP-filtered auxiliary data; inference includes cropping and additional enhancement procedures.
- Data Composition: The training data combine NTHazy target-domain pairs with CLIP-filtered samples from I-HAZE, Dense-Haze, and HAZE1K, totaling 59 paired images.
- Evaluation Protocol: NHM-20 contains 20 aligned image pairs distributed across five haze levels, enabling controlled public evaluation with Y-channel PSNR and SSIM as primary metrics.RGB-channel metrics and LPIPS are reported as additional references.
- Quantitative Results: The pipeline improves Y-channel PSNR and SSIM over the hazy-input baseline, but RGB metrics and LPIPS do not surpass the input baseline.The results indicate stronger brightness-structure recovery than color-fidelity or perceptual improvement.
- Ensemble Analysis: The 100K checkpoint achieves the best PSNR-Y, the 80K checkpoint achieves the best SSIM-Y, and weighted fusion stabilizes quality without optimizing every metric.Adaptive checkpoint selection may therefore offer further improvement.
- Overall Findings: NHM-20 confirms stable brightness-structure recovery, while perceptual improvement remains scene-dependent.
5. Conclusion
The paper concludes that a unified pipeline combining domain-consistent data screening, stage-wise training, and inference-time enhancement addresses data scarcity and prediction instability in nighttime dehazing. Public evaluation shows stable Y-channel improvements over hazy inputs, while RGB and LPIPS gains remain limited.
- The framework integrates CLIP-guided filtering, two-stage optimization, TLC-style local conversion, ×8 self-ensemble, and weighted snapshot fusion.
- NHM-20 evaluation confirms stable Y-channel PSNR and SSIM improvement over the hazy-input baseline.
- RGB metrics and LPIPS do not exceed the baseline, indicating that the main benefit lies in brightness-structure recovery rather than uniform perceptual improvement.
- Future work may refine cross-domain sample screening and degradation modeling closer to real nighttime imaging.