ControlLight: Towards Controllable, Consistent, and Generalizable Low-Light Enhancement

Yufeng Yang, Jianzhuang Liu, Jisheng Chu, Yuqi Peng, Xianfang Zeng, Jiancheng Huang, Shifeng Chen

arXiv:2605.25569 · 2026-05-26 공개 · arXiv · PDF

generalization image-restoration visual-consistency low-light-enhancement illumination-strength controllability control-light flow-matching-loss

Abstract

Existing deep learning-based low-light enhancement methods are typically trained on limited datasets with single enhancement targets, which restricts their generalization ability and controllability in real-world applications. To overcome these limitations, we propose ControlLight, a controllable, consistent, and generalizable framework for low-light enhancement. We first construct a large-scale dataset of real-world degraded images with continuous illumination-strength supervision. To further ensure consistent outputs under different control strengths, we introduce a misalignment-aware weighted flow matching loss that preserves image structure across continuous enhancement strengths. ControlLight allows users to edit real-world degraded low-light images toward satisfactory enhancement results by flexibly controlling the strength while preserving visual consistency and realism. Extensive experiments show that ControlLight achieves state-of-the-art performance against existing low-light enhancement approaches while demonstrating strong continuous controllability and generalization to real-world scenarios.

한국어 요약

📋 한 줄 요약

**[Low-Light Enhancement / Flow Matching]** ControlLight가 continuous illumination-strength supervision의 real-world degraded image 대규모 데이터셋과 misalignment-aware weighted flow matching loss로 일관된 강도 제어 가능 enhancement 달성, 실세계 일반화·controllability SOTA.

🎯 핵심 기여도

💡 핵심 아이디어

Low-light enhancement의 controllability·일관성은 single target 학습이 아니라 continuous illumination-strength supervision의 real-world 데이터셋 + flow matching의 strength 전반 structure 보존 loss로 달성되며, 사용자가 enhancement 강도를 연속 조절하면서도 일관된 visual·realism을 유지할 수 있다.

🔬 기술적 접근법

📊 주요 결과

💭 의의 및 한계

**의의**: Low-light enhancement의 controllability·일관성·일반화 3축을 단일 프레임워크에 통합, continuous strength supervision 대규모 데이터셋 자체로 후속 연구의 자산, flow matching loss의 misalignment-aware 변형이 일반 image-to-image task에 응용 가능. **한계**: 데이터셋 구축의 cost, continuous strength label 품질에 성능 의존, 극저조도·motion blur 결합 등 복잡 degradation에서의 robustness는 추가 검증 여지.

🚀 실용적 활용