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

AI summary

Overview

• Research area: Computational photography and generative computer vision, specifically low-light image enhancement (LLIE) using diffusion and flow-matching models.
• Technical level: Advanced. The paper assumes familiarity with flow matching, velocity fields, latent autoencoders, and diffusion-model acceleration techniques.
• Scope: A one-step latent flow-matching framework (GLOW-FM) that enhances low-light images with a single network evaluation using Terminal Velocity Matching in the latent space of a frozen Tiny-VAE, evaluated on real-world paired and unpaired low-light benchmarks.

What This Paper Is About

Low-light photos suffer from noise, color distortion, lost detail, and underexposure, and restoring them is ambiguous because information is genuinely missing. Generative models such as diffusion produce realistic results but need many iterative sampling steps, which is expensive. This paper asks whether a generative model can deliver high-quality low-light enhancement in a single step and within a small computational budget, while still generalizing to new cameras and lighting conditions.

Key Contributions

1. A one-step latent generative framework for LLIE. GLOW-FM performs conditional transport in the compact latent space of a frozen Tiny-VAE (TAESD) rather than pixel space, so an enhancement requires only a single function evaluation.
2. First adaptation of Terminal Velocity Matching (TVM) to image restoration. The authors apply TVM, which learns finite displacements between arbitrary time steps instead of integrating an instantaneous velocity field, and they report it as the first use of TVM for image restoration and specifically for LLIE.
3. Matched comparison of one-step objectives. Under identical settings, TVM is compared against standard Flow Matching, Consistency Flow Matching, and MeanFlow, supporting TVM as the strongest overall single-step objective for this task.
4. Strong cross-dataset generalization with low cost. The method reports state-of-the-art cross-dataset generalization against generative and non-generative baselines while staying competitive on image quality, with the small variant requiring 34 GFLOPs.

Main Findings

• Single-step quality on LSD-DLL: GLOW-FM-XL reaches a PSNR of 20.3, SSIM of 0.548, LPIPS of 0.48, MUSIQ of 36.0, and NIQE of 2.92 with 1 NFE, 48 GFLOPs, and 124.0M parameters. The small variant (GLOW-FM-S) obtains a PSNR of 19.7, LPIPS of 0.52, MUSIQ of 37.1, and NIQE of 2.70 at 34 GFLOPs and 9.1M parameters.
• Efficiency contrast with prior generative methods: WCDM requires 10 NFEs and about 90 GFLOPs, ReDDiT requires 2 NFEs but roughly 128 GFLOPs, and PyDiff requires 4 NFEs and 473 GFLOPs. GLOW-FM operates at 1 NFE with 34 to 48 GFLOPs across its variants.
• High-resolution benchmark: On UHD-LOL4K, GLOW-FM-XL reports PSNR 31.7, SSIM 0.969, LPIPS 0.12, MUSIQ 35.2, and NIQE 4.59. The paper notes that non-generative methods remain stronger on some distortion-oriented metrics there, while GLOW-FM keeps competitive quality at much lower inference cost than generative baselines.
• Generalization to unseen devices: On the LSD-U split, GLOW-FM achieves the best NIQE and MUSIQ among all compared generative and non-generative methods with a single NFE (XL: NIQE 4.33, MUSIQ 28.4; M: 4.49 / 28.3; S: 4.99 / 27.7).
• Cross-dataset generalization: Trained on LOL-v2-Syn and tested on DICM, LIME, MEF, NPE, and VV without fine-tuning, GLOW-FM variants obtain mean NIQE of 2.86 (XL), 2.84 (M), and 2.83 (S) with mean BRISQUE of 10.8, 11.1, and 10.9, versus the best baseline means of 3.29 NIQE (WCDM) and 13.8 BRISQUE (Reti-Diff).
• LOL-v1 trade-off: GLOW-FM achieves the best NIQE on LOL-v1 with a single NFE (XL 3.71, M 3.79, S 3.70) but weaker distortion performance (XL PSNR 21.8), which the authors attribute to limited training data in that benchmark.
• Ablation on one-step objectives: With architecture and settings fixed on LSD-DLL, TVM (PSNR 20.3, MUSIQ 36.0, NIQE 2.92) outperforms ConsistencyFM (19.4 / 32.8 / 3.53) and MeanFlow (19.5 / 31.7 / 3.59), and also outperforms standard Flow Matching at 1 NFE (19.7 / 27.7 / 6.00) and at 25 NFEs (20.1 / 35.6 / 3.00).
• Latent space matters: A pixel-space TVM variant needs 1693 GFLOPs and 121M parameters yet achieves only PSNR 16.0, MUSIQ 28.2, and NIQE 4.51. SD-VAE latents cost 471 GFLOPs and 205M parameters and give PSNR 19.6, LPIPS 0.47, MUSIQ 40.6, and NIQE 3.37, while TAESD gives better PSNR, SSIM, and NIQE at far lower cost.
• Autoencoder efficiency: Replacing SD-VAE with TAESD reduces encoding/decoding GFLOPs by 15x and parameter count by 33x, with minimal loss in end-to-end restoration performance.
• Data scaling: Training on progressively larger LSD-DLL subsets from approximately 0.2K to 3.0K image pairs improves PSNR monotonically from 16.8 to 20.3 dB and MUSIQ from 34.5 to 36.0.

Methodology in Plain English

The paper treats enhancement as a transport problem: the model must move a low-light image to its well-exposed counterpart. Instead of doing this in pixel space, both images are first compressed into a compact latent representation using a frozen, lightweight Tiny-VAE (TAESD), and only the flow model is trained. Following TVM's time convention, the low-light latent sits at time t = 1 and the normal-light latent at t = 0, so enhancement moves toward earlier times.

The core network is a two-time conditioned model built on an Hourglass Diffusion Transformer (HDiT) backbone, which limits expensive global attention to low-resolution stages and uses windowed attention elsewhere, giving linear rather than quadratic scaling. Training combines two losses: a standard flow-matching loss that anchors the instantaneous velocity at the boundary where start and end times coincide, and a terminal-velocity loss that supervises the finite displacement between two sampled times. Because the true terminal velocity is unavailable, an EMA copy of the network supplies the target, with a stop-gradient, and the derivative with respect to the terminal time is computed with a Jacobian-vector product. A weighting term for the terminal-velocity loss is warm-started, so the velocity field is established before the displacement objective is introduced. During training the low-light latent is perturbed with small Gaussian noise (σ = 0.01) to smooth the source distribution.

At inference, the low-light image is encoded once, and the model predicts the displacement over a uniform time grid; the number of intervals equals the number of function evaluations. With M = 1, a single network call transports the latent directly to the enhanced endpoint, and the result is decoded back to an image. The same displacement model also supports a few-step refinement mode. The authors additionally derive a bound on the pixel-space Wasserstein-2 distance between restored and target distributions in terms of the TVM objectives and autoencoder reconstruction error, under TVM regularity conditions and a Lipschitz assumption on the decoder (Appendix B). Models are trained for roughly 100k optimization steps on 8 NVIDIA H100 GPUs, processing 16-channel latents, and results are averaged over five evaluation seeds.

Why This Matters

Impact on research. The paper makes the case that generative priors for image restoration need not come with iterative sampling costs. It also introduces TVM into image restoration and provides a matched comparison against several one-step alternatives, giving the field a concrete data point on which single-step objective works best for LLIE and why.

Real-world applications:

• Smartphone camera pipelines, where per-shot compute and memory budgets are tight and multi-step diffusion is impractical.
• Surveillance and automotive imaging, where poorly exposed frames degrade downstream detection and recognition.
• Consumer and drone photography needing on-device enhancement of high-resolution images such as the UHD-LOL4K benchmark.
• Image processing in resource-constrained or battery-powered devices, where the small variant's 34 GFLOPs and 9.1M parameters matter.

Industry relevance. The work comes from Arm AI Research (Arm Ltd), and its emphasis on single-NFE inference, compact latents, and reduced autoencoder cost reflects deployment constraints on edge hardware. The scale of the reported savings, such as a 15x reduction in encoding/decoding GFLOPs and 33x fewer parameters from swapping SD-VAE for TAESD, is directly relevant to shipping generative enhancement on device.

Future Directions

• Closing the distortion gap under small data. The weaker LOL-v1 distortion results are attributed to limited training data, supported by a data-scaling experiment, but how GLOW-FM behaves with genuinely scarce paired data is left open.
• Balancing latent choice. TAESD gives better PSNR, SSIM, and NIQE at low cost, but SD-VAE gives better LPIPS (0.47) and MUSIQ (40.6) at much higher cost; finding a latent that wins on both axes is an open design question.
• Extending beyond single-step objectives. The paper compares TVM with Rectified Flow, Consistency Flow Matching, and MeanFlow; whether hybrid objectives or few-step schedules can improve the perception-distortion balance further is not resolved.
• Broadening restoration scope. Since this is described as the first adaptation of TVM to image restoration, applying the same one-step latent formulation to other degradation types, such as deblurring or denoising, is a natural next step. The paper does not report such experiments.

Target Audience

Researchers and engineers working on generative image restoration, diffusion and flow-matching acceleration, and on-device computational photography. It is most useful to readers with a working knowledge of flow matching and latent autoencoders who are evaluating whether single-step generative models can meet real deployment budgets; practitioners focused on mobile or edge camera pipelines will find the efficiency and generalization results particularly relevant.

Authors’ 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.