Low-Light Photos in One Step: GLOW-FM Wants to Bypass Diffusion Sampling
Arm’s GLOW-FM puts low-light enhancement into the latent space of a frozen Tiny-VAE and uses TVM for one-step generation. The small variant is 34 GFLOPs and achieves SOTA cross-dataset generalization; however, its PSNR on LOL-v1 still trails WCDM. Real-time on phones remains to be seen in actual tests.
The problems with low-light photos are very concrete: noisy shadows, color casts, and details smeared into a blur. Brightening is easy; making it believable is hard. Diffusion models have raised the quality of low-light enhancement, but they have also raised sampling costs. GLOW-FM, from an Arm research team, aims to solve the latter: reduce generative low-light image enhancement to a single step.
First, the demo:
The video comes from X post. In the footage you can see night-time buildings and reflections on water, as well as the colors and details of flowers and green leaves being brought back. The demo is intuitive, but this kind of side-by-side video only shows one thing: the method can produce results; it does not show how fast it can run on a phone.
The paper title is GLOW-FM: Efficient Generative Low-Light Image Enhancement with One-Step Latent Flow Matching. The method is not complicated: first, feed the low-light image into a frozen Tiny-VAE and compress it into a compact latent space; then restore it in latent space with an Hourglass Diffusion Transformer (HDiT); train with Terminal Velocity Matching (TVM), so enhancement is completed in a single function evaluation. The authors claim this is the first adaptation of TVM to image restoration, and the first for low-light enhancement specifically.
There are two key metrics in this design. One is single-step inference, which the paper calls single function evaluation; the other is compute, with the small variant requiring only 34 GFLOPs. Compared with diffusion-based generative methods that repeatedly sample, these two numbers point directly at deployment cost. The paper also claims GLOW-FM achieves competitive image quality and SOTA cross-dataset generalization on real-world benchmarks. The project page, paper, and GitHub are all up, and the paper says code will be released after publication.
But don’t rush to equate 'one-step' with 'winning across the board.' A commenter on X pointed out that on LOL-v1, GLOW-FM’s PSNR is 21.8 dB, lower than WCDM’s 26.3 dB. The same reply mentioned that the Arm team attributes the gap to data: after increasing training pairs from 0.2K to 3.0K, PSNR on LSD-DLL rose from 16.8 dB to 20.3 dB. This shows that generative low-light enhancement still involves a trade-off between perceptual quality and pixel fidelity, and data scale can directly change results. Strong cross-dataset generalization does not mean every benchmark’s PSNR looks good.
The comments also raise more practical questions. Some ask whether this can run in real time on phone chips; some say food photos taken in dim restaurants finally get a second chance; others mention blurry phone photos taken at 2 a.m. These scenarios illustrate the need for low-light enhancement better than paper tables do: don’t let shadow detail and color fall apart, and don’t turn night scenes into daytime. That is also where GLOW-FM’s value lies: one step, a compact latent space, and a frozen VAE offer a cheaper deployment path for generative low-light enhancement.
Whether it can really get into phones will depend on the follow-up code, on-device measurements, and stability under different low-light conditions. The flowers in the demo video look great, but the deployment ledger will depend on how 34 GFLOPs performs on real chips.
Paper: https://huggingface.co/papers/2610.31222
Project page: https://arm-research.github.io/AAIR-GLOW-FM
GitHub: https://github.com/arm-research/AAIR-GLOW-FM
发布时间: 2026-11-03 05:42