CLIP-Guided Data Augmentation for Night-Time Image Dehazing

TL;DR

This paper introduces a practical framework for nighttime image dehazing that leverages CLIP-guided data selection, stage-wise training, and inference enhancements to address domain mismatch and improve stability.

Contribution

It presents a unified approach combining CLIP-based sample selection, staged training, and inference-time techniques without complex network redesign for effective nighttime dehazing.

Findings

CLIP-guided data construction improves domain relevance of training samples.

Stage-wise training enhances adaptation to target domain degradation.

Inference-time ensemble and fusion improve output stability.

Abstract

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.