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

GLOW-FM. Efficient Generative Low-Light Image Enhancement with One-Step Latent Flow Matching. By Ariel Lapid and four others, from Arm AI Research. Published on the twenty-second of October.

Low-light image enhancement tries to recover images that are well exposed and perceptually faithful from observations that are severely degraded. Diffusion-based generative methods reach strong perceptual quality by drawing on powerful image priors, but their sampling is iterative, and that computational cost limits practical deployment. The priors that make the results look right are reached through repeated evaluations, and every one of them has to be paid for. Generative enhancement has recently become more efficient. Even so, high-quality results from a single function evaluation, at low overall computation, remain hard to achieve, particularly across the diverse low-light conditions of the real world.

GLOW-FM is a one-step latent flow-matching framework built for exactly that gap. It performs conditional transport inside the latent space of a frozen Tiny-VAE, and it relies on Terminal Velocity Matching, or TVM, to make enhancement possible with a single function evaluation. The autoencoder stays frozen, so what is learned is the transport within its latent space, not the space itself.

Two levers act on cost here. Working in a compact latent space keeps each evaluation cheap, and finishing in one step means there is only one of them. The choice of TVM is tested rather than assumed: matched comparisons against alternative one-step objectives support it as an effective design choice. To the authors' knowledge, this is also the first time TVM has been adapted to image restoration, and specifically to low-light enhancement.

Extensive experiments on challenging real-world benchmarks report competitive image quality, together with state-of-the-art cross-dataset generalization. Those results come at a single function evaluation, and the small variant needs only thirty-four GFLOPs. Generalization across datasets matters for this task in particular, since the difficulty named at the outset was the variety of real low-light conditions. A method tuned to one collection of dark images and then tested on another is being asked exactly that question.

The authors frame the result as moving generative enhancement closer to practical deployment, through single-step inference and compact latent modelling. Diffusion pays for its priors with many sampling steps. GLOW-FM aims to keep the generative approach while paying for it once, and the cross-dataset results are the evidence offered that one step is enough.