IllumFlow: Illumination-Adaptive Low-Light Enhancement via Conditional Rectified Flow and Retinex Decomposition

TL;DR

IllumFlow tackles low-light image enhancement by decomposing an image as $I = L \odot R$ and applying a conditional rectified flow (CRF) to model illumination changes as a continuous flow, while a flow-augmented denoiser improves reflectance. The illumination flow supports bidirectional and continuous brightness adjustments, enabling explicit exposure control, and the reflectance denoiser uses data augmentation guided by flow to suppress noise and chromatic aberrations. The method combines a pretrained Retinex decomposition network (TDN from Diff-Retinex), a CRF illumination module, and a CRFR denoiser, achieving superior results on LOLv1/LOLv2/MEF with fast, single-inference illumination control. The reported gains over end-to-end, diffusion-based, and Retinex-based LLIE methods highlight improved color fidelity, noise suppression, and generalization to diverse illumination conditions.

Abstract

We present IllumFlow, a novel framework that synergizes conditional Rectified Flow (CRF) with Retinex theory for low-light image enhancement (LLIE). Our model addresses low-light enhancement through separate optimization of illumination and reflectance components, effectively handling both lighting variations and noise. Specifically, we first decompose an input image into reflectance and illumination components following Retinex theory. To model the wide dynamic range of illumination variations in low-light images, we propose a conditional rectified flow framework that represents illumination changes as a continuous flow field. While complex noise primarily resides in the reflectance component, we introduce a denoising network, enhanced by flow-derived data augmentation, to remove reflectance noise and chromatic aberration while preserving color fidelity. IllumFlow enables precise illumination adaptation across lighting conditions while naturally supporting customizable brightness enhancement. Extensive experiments on low-light enhancement and exposure correction demonstrate superior quantitative and qualitative performance over existing methods.

Figures (20)

• Figure 1: Comparison of SWANet 10244055 and our method outputs for input images under different lighting conditions on datasetsafifi2021learning originally rendered from the MIT-Adobe FiveK dataset.
• Figure 2: Comparison with learning- and diffusion-based methods: (a-c) RetinexNet wei2018deep shows artifacts/over-enhancement (top) vs. our method's cleaner results (bottom); (d-e) DiffLL jiang2023low produces unnatural colors (left) while our method maintains better color fidelity and illumination (right).
• Figure 3: Training process. This framework involves: (1) TDN-based image decomposition to separate illumination/reflectance layers; (2) Enhanced denoising of reflectance; (3) Flow-based (CRF) continuous illumination enhancement for smooth brightness adjustment.
• Figure 4: Variations in pixel values at fixed positions across ten representative images as a function of exposure time.
• Figure 5: Inference process. Color-coded framework components: Green: TDN decomposition (preprocessing); Pink: Enhanced denoising module; Purple: Flow-based illumination enhancement (progressive process; single CRFI = one-step enhancement); Yellow: Final output (denoised reflectance + enhanced illumination sequence).