Abstract
Low-light image enhancement (LLIE) models face severe performance degradation when deployed from easy-to-access paired daily scenes to specialized domains like medical endoscopy and remote sensing, where paired supervision is unavailable. While existing solutions rely on scarce target-side labels or heuristic unsupervised priors, we propose COLD, a COunterfactual and Lighting Debiasing approach for source-free domain adaptation in low-light enhancement. COLD attributes the performance drop to observable lighting bias and unobservable domain bias, addressing them through two novel modules within a teacher-student framework. Specifically, a lighting debiasing module dynamically aligns student outputs with normal-exposure references and low-uncertainty teacher predictions to stabilize illumination recovery. Simultaneously, a counterfactual-inspired debiasing module probes source-related bias by feeding content-free inputs into the teacher and treating the teacher's content-independent responses as negative anchors, sub-sequently distancing student features from these biased anchors through contrastive learning. Experiments on diverse, challenging cross-domain transfers demonstrate that COLD improves the source model by up to 11 dB in peak signal-to-noise ratio, outper-forming state-of-the-art unsupervised methods by 5 dB. These results validate that COLD effectively bridges the domain gap through empirically grounded bias mitigation, ensuring robust performance in specialized low-light environments. The code is available at COLD .