Incorporating DINO Priors into Flow Matching for Low-Light Image Enhancement

Abstract

Flow matching enables efficient low-light image enhancement (LLIE) with very few sampling steps, yet standard architectures lack explicit scene understanding, causing structural degradation and artifacts in challenging regions. We propose DINO-guided Flow Matching, which leverages a frozen DINOv3 backbone to provide illumination and structure priors for the Pixel MeanFlows framework. Specifically, we extract dual-layer features—shallow illumination-sensitive features and deep degradation-invariant structure features—and bridge the low-light/normal-light domain gap through a lightweight DINO Feature Corrector (DFC). The corrected features are injected into the flow-matching UNet via Retinex-inspired FiLM modulation and cross-attention, providing spatially adaptive guidance. Furthermore, we identify a systematic brightness drift problem arising from the marginal distribution mismatch between source and target domains, and address it with an Optimal-Transport Look-Up Table (OT-LUT) that pre-aligns the intensity distribution at negligible cost. Experiments on LOL-v2-real, LOL-v2-synthetic, and MIT-5K demonstrate state-of-the-art results in both distortion metrics (PSNR, SSIM) and perceptual quality (LPIPS).