GLOW-FM | Arm Research

TL;DR

Generative low-light image enhancement using Flow Matching

with only

1NFE

34GFLOPs

9.1Mparameters

GFLOPs and parameter counts are for the small variant, GLOW-FM-S.

Kitchen sink scene: Input. — Low-light input, GLOW-FM enhancement, and ground truth (GT) on LSD-DLL. Hover over the input to explore details.

Kitchen sink scene: GLOW-FM. — Low-light input, GLOW-FM enhancement, and ground truth (GT) on LSD-DLL. Hover over the input to explore details.

The idea behind GLOW-FM

GLOW-FM combines a lightweight Tiny-VAE encoder with an Hourglass Diffusion Transformer (HDiT) architecture to perform efficient enhancement in a compact latent space. Trained with Terminal Velocity Matching (TVM), the HDiT restores low-light latents in a single step. Together, the compact architecture and single-step inference enable efficient generative low-light image enhancement.

EfficiencyEfficient generative enhancement. Compact latent flow matching delivers one-step restoration at low computational cost.
Terminal Velocity MatchingTVM for image restoration. Adapted to low-light enhancement and validated against alternative one-step objectives in matched comparisons.
GeneralizationStrong generalization. State-of-the-art cross-dataset results among the compared methods, with competitive image quality.

Figure 1 radar chart comparing GLOW-FM with PyDiff, WCDM, and ReDDiT across image quality, efficiency, and generalization. Farther from the center is better. — **Figure 1. GLOW-FM Performance.** Comparison with efficient generative LLIE baselines on LSD-DLL. Metrics are min–max normalized across the compared methods, with lower-is-better metrics reversed so that farther from the center is always better; axis annotations show the corresponding absolute metric values. *Perception Generalization* uses LSD-U NIQE.

Abstract

Low-light image enhancement (LLIE) aims to recover well-exposed and perceptually faithful images from severely degraded observations. Recent diffusion-based generative methods achieve strong perceptual quality by leveraging powerful image priors, but their iterative sampling incurs substantial computational cost, limiting practical deployment. Despite recent efficiency gains in generative LLIE, high-quality enhancement with a single function evaluation and low overall computation remains challenging, particularly under diverse real-world low-light conditions.

We introduce GLOW-FM, a one-step latent flow-matching framework for efficient generative LLIE. GLOW-FM performs conditional transport in the latent space of a frozen Tiny-VAE and leverages Terminal Velocity Matching (TVM) to enable enhancement with a single function evaluation. Matched comparisons with alternative one-step objectives support TVM as an effective design choice, and, to the best of our knowledge, this is its first adaptation to image restoration and specifically to LLIE.

Extensive experiments on challenging real-world benchmarks demonstrate that GLOW-FM achieves competitive image quality and state-of-the-art cross-dataset generalization while requiring only a single function evaluation and 34 GFLOPs for the small variant. These findings highlight the potential of GLOW-FM to bring generative LLIE closer to practical deployment through single-step inference and compact latent modeling. Code will be made publicly available upon publication.

How it works

Method

GLOW-FM training diagram: paired images enter a frozen encoder, followed by flow-matching and terminal-velocity-matching objectives with an EMA teacher. — **Figure 2.** GLOW-FM training scheme.

01 Construct the latent path

Encode paired images with a frozen Tiny-VAE encoder. Perturb the low-light latent with Gaussian noise, then interpolate toward the normal-light latent.

\[z_t=(1-t)z_0+t\tilde z_1\]

02 Learn the displacement

HDiT parameterizes the two-time displacement model \(F_\theta(z_t,t,s)\), conditioned on start and end times. At \(t=s\), \(F_\theta\) is supervised as instantaneous velocity; multiplying by \(s-t\) gives the displacement \(f_\theta\).

\[f_\theta(z_t,t,s)=(s-t)F_\theta(z_t,t,s)\]

03 Match terminal velocity

TVM matches \(\partial_s f_\theta\), computed via JVP, to the stop-gradient EMA teacher’s velocity at the predicted endpoint.

\[\mathcal L=\mathcal L_{\mathrm{FM}}+\lambda\mathcal L_{\mathrm{TVM}}\]

Inference · one network evaluation

Encode → transport directly from \(t=1\) to \(s=0\) → decode.

\[\hat x_0=\mathcal D\!\left(z_1+f_\theta(z_1,1,0)\right),\qquad z_1=\mathcal E(x_1)\]

See the differences across methods and datasets

Qualitative examples

Compare GLOW-FM with ReDDiT, PyDiff, and WCDM on LSD-U and LSD-DLL, UHD-LOL4K, LOL-v1, and the generalization datasets (MEF, VV, NPE, LIME, and DICM). Ground truth is shown where available.

LSD-U Scene 2 — Input — **LSD-U** · Example 1

LSD-U Scene 2 — ReDDiT — **LSD-U** · Example 1

Hover or tap any image to zoom into the same region across methods. Use the arrows to browse scenes.

Quality and efficiency

Quantitative results

Three model sizes offer different quality and compute tradeoffs. All GLOW-FM variants use one function evaluation.

LSD-DLL · Generative methods from Tab. 1 in the paper
Method	NFEs ↓	GFLOPs ↓	Params (M) ↓	PSNR ↑	SSIM ↑	LPIPS ↓	MUSIQ ↑	NIQE ↓
PyDiff	4	473	97.9	17.3	0.544	0.56	33.2	7.38
WCDM	10	90	22.1	17.1	0.553	0.59	25.5	5.94
ReDDiT	2	128	17.4	14.1	0.488	0.59	30.9	7.45
GLOW-FM-XL	1	48	124	20.3	0.548	0.48	36.0	2.92
GLOW-FM-M	1	36	21.8	19.5	0.531	0.50	35.5	2.86
GLOW-FM-S	1	34	9.1	19.7	0.516	0.52	37.1	2.70

↑ Higher is better. ↓ Lower is better. Bold indicates the best value among the methods shown, including ties. This excerpt compares generative methods; see the paper for non-generative baselines, other datasets, and evaluation details.

Additional results

Unseen devices

GLOW-FM-XL achieves 4.33 NIQE and 28.4 MUSIQ on LSD-U, the unseen-device split (Tab. 2 in the paper).

Cross-dataset generalization

GLOW-FM-S achieves a mean NIQE of 2.83 across DICM, LIME, MEF, NPE, and VV, the lowest among methods compared in Tab. 4 in the paper.

Measured inference speed

GLOW-FM-S runs in 53.8 ms at 1024 × 1024 on an NVIDIA H100 (Tab. 15 in the paper).

Build on this work

Citation

If you find this work useful, please consider citing our paper.

@misc{lapid2026glowfmefficientgenerativelowlight,
      title={GLOW-FM: Efficient Generative Low-Light Image Enhancement with One-Step Latent Flow Matching},
      author={Ariel Lapid and Yuval Haitman and Reuven Peretz and Arnon Netzer and Idit Diamant},
      year={2026},
      eprint={2610.31222},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2610.31222},
}