GLOW-FM: Efficient Generative Low-Light Image Enhancement with One-Step Latent Flow Matching

GLOW-FM is a one-step latent flow-matching framework for low-light image enhancement that uses Terminal Velocity Matching in a frozen Tiny-VAE latent space and requires a single function evaluation.

Low-light images contain severe underexposure, noise, color distortion, reduced contrast, detail loss, and local saturation that vary across cameras and imaging conditions. Regression-based LLIE methods are efficient but often produce overly smooth results and suppress fine details under severe degradation. Diffusion-based generative LLIE improves perceptual quality, but iterative sampling is computationally costly. Recent efficient generative LLIE reduces NFEs, yet still relies on iterative refinement and substantial computation, leaving single-step high-quality enhancement challenging.

GLOW-FM formulates LLIE as conditional flow matching between low-light distribution p1(x1) and normal-light distribution p0(x0), with the source at t=1 and target at t=0. It performs transport in the latent space of a frozen VAE, using z1=E(x1) and z0=E(x0), and trains a two-time conditioned model Fθ(z_t,t,s) with Terminal Velocity Matching. The model parameterizes finite displacement as fθ(z_t,t,s)=(s-t)Fθ(z_t,t,s), recovering instantaneous velocity at s=t. Training combines an FM loss on the instantaneous velocity with a TVM loss that matches the derivative with respect to terminal time to an EMA target velocity, using stop-gradient and a warm-up schedule for λ. The source latent is perturbed with Gaussian noise σ=0.01; at inference, M=1 uses t=1 and s=0 to transport the encoded low-light latent in one step and decode it.

• GLOW-FM-S reaches 19.7 PSNR and 37.1 MUSIQ on LSD-DLL with 1 NFE and 34 GFLOPs. §4.2
• GLOW-FM-S attains mean NIQE 2.83 and BRISQUE 10.9 across DICM, LIME, MEF, NPE, and VV. §4.3
• On LSD-U, GLOW-FM-XL obtains NIQE 4.33 and MUSIQ 28.4 with 1 NFE. §4.3
• Among matched TAESD one-step variants, GLOW-FM with TVM obtains PSNR 20.3, SSIM 0.55, MUSIQ 36.0, and NIQE 2.92 at 48 GFLOPs. §4.4
• Increasing training-set size from approximately 0.2K to 3.0K image pairs improves PSNR from 16.8 to 20.3 dB and MUSIQ from 34.5 to 36.0. §4.4
• Replacing SD-VAE with TAESD reduces encoding/decoding GFLOPs by 15× and parameter count by 33×, with minimal loss in end-to-end restoration performance. §3.4

• No dedicated limitations section is provided. The paper reports weaker distortion performance on LOL-v1 and notes that non-generative methods remain stronger on some distortion-oriented metrics on UHD-LOL4K. [§4.2]
• Cross-dataset generalization is evaluated on unpaired datasets without paired references, using only no-reference NIQE and BRISQUE metrics. [§4.1]
• The LOL-v1 distortion weakness is hypothesized to come from limited training data; the paper supports this with a data-scaling experiment but does not show it resolves the gap on all benchmarks. [§4.2]

• The source is placed at t=1 and target at t=0, opposite the original Flow Matching convention. §3
• The two-time model predicts finite displacements, and scaling enforces zero displacement at s=t. §3.1
• The FM loss supervises the boundary case where s=t and Fθ represents instantaneous velocity. §3.2
• The TVM loss uses an EMA copy as target proxy and stop-gradient on the predicted terminal latent. §3.2
• Inference with M=1 uses t=1 and s=0 and needs one network evaluation. §3.3
• Source latents are perturbed with Gaussian noise using σ=0.01. §4.1
• HDiT restricts global attention to low-resolution stages and uses fixed-window neighborhood attention at higher resolutions. §3.4
• GLOW-FM-S uses 34 GFLOPs and 9.1M parameters with one NFE. §4.2