PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Ariel Lapid, Yuval Haitman, Reuven Peretz, Arnon Netzer, Idit Diamant

arXiv:2610.31222v1cs.CV

TL;DR

Low-light enhancement must recover faithful images despite variable degradations, while combining one-step inference, low computation, and robust real-world performance remains challenging. GLOW-FM performs conditional transport in a frozen Tiny-VAE latent space using TVM. It reports competitive image quality and state-of-the-art cross-dataset generalization with one function evaluation.

  • Problem

    Combining high-quality single-step enhancement, low computation, and robust performance across diverse real-world low-light conditions remains challenging.

    Source: §1 · p. 1, §5 · p. 8

  • Method

    GLOW-FM performs conditional transport in a frozen Tiny-VAE latent space and uses Terminal Velocity Matching for single-step enhancement.

    Source: §1 · p. 2, §3 · p. 3

  • Results

    GLOW-FM achieves competitive image quality and state-of-the-art cross-dataset generalization while requiring a single NFE.

    Source: §1 · p. 2, §5 · p. 8

  • Takeaways & Limitations

    One-step latent flow matching with TVM is presented as a practical direction for robust, efficient generative low-light enhancement.

    Source: §5 · p. 8

Abstract

from arXiv · show

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.

1. Introduction

GLOW-FM targets the challenge of combining high-quality, single-step enhancement with low computation across diverse real-world conditions. It uses one-step latent flow matching with TVM and reports competitive image quality and state-of-the-art cross-dataset generalization.

1. Introduction

  • Real-world low-light images combine variable degradations, and poor exposure also impairs downstream detection and recognition.The stated goal is perceptually faithful restoration that supports reliable high-level vision applications.Source: §1 · p. 1
  • High-quality enhancement with one inference step and low overall computation remains challenging across diverse, demanding real-world low-light conditions.WCDM uses 10 NFEs and approximately 90 GFLOPs, while ReDDiT uses 2 NFEs and roughly 128 GFLOPs.Source: §1 · p. 2
  • GLOW-FM uses TVM for direct normal-light prediction in the frozen Tiny-VAE latent space, avoiding iterative inference sampling.The paper presents matched comparisons with standard Flow Matching, Consistency Flow Matching, and MeanFlow as support for TVM as a one-step LLIE design choice.Source: §1 · p. 2
  • GLOW-FM achieves state-of-the-art cross-dataset generalization and competitive image quality with a single NFE.The authors present it as a favorable quality–efficiency trade-off for generative LLIE.Source: §1 · p. 2

2. Related Work

Prior LLIE work spans direct regression and diffusion-based restoration, with later methods reducing diffusion sampling costs. Flow-based acceleration methods have explored one- and few-step generation, but efficient flow-based generation remains relatively unexplored for LLIE.

2.1. Generative Low-Light Image Enhancement

  • LLIE research has moved from efficient direct regression toward diffusion-based methods that seek more realistic reconstructions, followed by efforts to reduce sampling cost.Examples include Retinex-based diffusion, structure-aware diffusion, latent Retinex priors, and wavelet-domain diffusion.Source: §2.1 · p. 2
  • GLOW-FM encodes paired low-light and normally exposed images into a frozen autoencoder latent space, learning velocity and terminal-displacement predictions for single-step inference.The enhanced latent is decoded back to image space.Source: Figure 2 · p. 3

2.2. Generative Model Acceleration

  • Flow-based acceleration includes straight-trajectory, velocity-consistency, mean-trajectory, and finite-displacement objectives, but efficient flow generation remains relatively unexplored for LLIE.The related-work discussion positions GLOW-FM as a single-step latent flow-matching framework for low-light enhancement.Source: §2.2 · p. 3

3. Method

GLOW-FM transports low-light inputs toward normal-light targets in a frozen VAE’s latent space, using Terminal Velocity Matching (TVM) to support one-step enhancement. Its training and architecture combine flow and TVM objectives with compact latent modeling for efficient inference.

3. Method

  • GLOW-FM performs conditional transport from low-light to normal-light images in a frozen VAE’s latent space, using a two-time TVM model for one-step enhancement or few-step refinement.The method also derives a pixel-space Wasserstein-2 bound under TVM regularity conditions and a Lipschitz decoder assumption, incorporating autoencoder reconstruction error.Source: §3 · p. 3

3.1. Terminal Velocity Matching

  • TVM parameterizes finite displacements between times, using terminal velocity as a practical proxy to supervise them without explicit ODE integration.The model’s two-time parameterization enforces zero displacement at s = t, where it recovers instantaneous velocity.Source: §3.1 · p. 4

3.2. Training

  • Training perturbs the low-light source latent with Gaussian noise, then combines flow-matching and TVM losses to supervise latent transport.An EMA model supplies a stop-gradient terminal-velocity target computed with a Jacobian-vector product; the TVM weight is gradually introduced after initial flow-matching training.Source: §3.2 · pp. 4–5

3.3. Inference

  • At inference, each time interval requires one network evaluation; setting M = 1 directly transports the encoded low-light latent from t = 1 to t = 0.The same displacement model can also be applied over multiple shorter intervals.Source: §3.3 · p. 5

3.4. Efficient Latent Architecture

  • Compact latent processing lowers the cost of each function evaluation, while the lightweight TAESD autoencoder reduces encoding and decoding computation.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. The HDiT backbone confines global attention to low-resolution stages and uses fixed-window neighborhood attention at higher resolutions for linear O(n) scaling.Source: §3.4 · pp. 5–6

4. Experiments

Across paired and unpaired real-world benchmarks, GLOW-FM balances enhancement quality with low inference cost and generalizes to unseen devices and domains. Ablations examine its one-step objective, training-set size, and latent representation.

4.1. Experimental Setup

  • Evaluation covers paired LSD-DLL, UHD-LOL4K, and LOL-v1, plus unpaired unseen-device and cross-dataset benchmarks.Paired benchmarks use PSNR, SSIM, and LPIPS alongside NIQE and MUSIQ; unpaired benchmarks use NIQE and BRISQUE.Source: §4.1 · pp. 6–7

4.2. Results on Paired Benchmarks

  • Across paired LSD-DLL and UHD-LOL4K benchmarks, GLOW-FM offers a favorable quality–efficiency trade-off with one NFE.It leads across several distortion and perceptual metrics on LSD-DLL; on UHD-LOL4K, non-generative methods remain stronger on some distortion metrics, while GLOW-FM variants reduce computation and retain strong quality.Source: §4.2 · p. 7

4.3. Generalization to Unseen Low-Light Data

  • On five unpaired real-world benchmarks, GLOW-FM achieves the best generative NIQE and BRISQUE results on nearly all datasets and the strongest mean across both metrics.It is evaluated directly after training on LOL-v2-Syn, without target-dataset fine-tuning.Source: §4.3 · p. 7, Table 4 · p. 7
  • On unseen-device LSD-U images, GLOW-FM achieves the best NIQE and MUSIQ scores among compared methods with a single NFE.All variants retain competitive no-reference image quality at substantially lower cost than competing generative methods.Source: §4.3 · p. 7
  • Unpaired examples show natural illumination and fewer artifacts, complementing paired ground-truth comparisons of the perception–distortion balance.The paper also reports additional qualitative examples in its appendix.Source: §4.3 · p. 8, Figure 4 · p. 8

4.4. Ablation Study

  • With training size increased from approximately 0.2K to 3.0K LSD-DLL pairs, PSNR rises from 16.8 to 20.3dB and MUSIQ from 34.5 to 36.0.Architecture and training protocol are held fixed.Source: §4.4 · p. 8, Figure 5
  • Among matched TAESD one-step variants, TVM achieves the strongest overall distortion and perceptual performance.The comparison includes MeanFlow, ConsistencyFM, and standard Flow Matching; No-TVM is also evaluated at 25 NFEs.Source: §4.4 · p. 8, Table 5 · p. 8
  • TAESD improves PSNR, SSIM, and NIQE over SD-VAE with nearly an order of magnitude less computation, while SD-VAE improves LPIPS and MUSIQ at higher cost.Direct pixel-space modeling costs significantly more and produces weaker distortion and perceptual quality than the latent variant.Source: §4.4 · p. 8, Table 5 · p. 8

5. Conclusion

GLOW-FM addresses the challenge of achieving one-step, low-computation generative enhancement that remains robust across real-world conditions. Its experiments report competitive quality, improved efficiency, and state-of-the-art cross-dataset generalization.

  • Real-world low-light enhancement remains ambiguous under noise, color shifts, and information loss, while combining single-step inference, low computation, and robust performance remains challenging.Source: §5 · p. 8
  • GLOW-FM combines one-step latent flow matching with Terminal Velocity Matching for single-NFE enhancement while retaining generative priors for realistic restoration.Source: §5 · p. 8
  • Experiments report competitive quality, improved efficiency, and state-of-the-art cross-dataset generalization for GLOW-FM.Source: §5 · p. 8

Latent Flow Matching

The supplementary material collects additional theoretical, implementation, training, ablation, quantitative, and qualitative material supporting the main paper.

B. Pixel-Space Wasserstein Bound for Latent TVM

The appendix relates latent TVM training to a pixel-space Wasserstein-2 gap between predicted and true normal-light image distributions. The bound depends on the latent training objective and the frozen VAE’s reconstruction error.

C. Additional Implementation Details

The implementation uses a frozen Tiny-VAE and trains only the latent flow model. It specifies training schedules, input crops, and a gradual TVM-weight schedule for LSD-DLL and UHD-LOL4K.

Training and inference

  • Unless otherwise specified, experiments optimize only the latent flow model with a frozen Tiny-VAE encoder/decoder on 8 NVIDIA H100 GPUs, using effective batch size 16 for approximately 100k optimization steps, AdamW, and learning rate 10−5.Gradient clipping is at norm 1.0.Source: C. Additional Implementation Details · p. 14
  • 2048×2048 crops are used for LSD-DLL and UHD-LOL4K, while LOL-v1 uses 384×384 crops; evaluation uses original image resolution.Source latents receive Gaussian noise with σ = 0.01 during training and σ = 0 during evaluation; reported results use one NFE unless specified otherwise.Source: C. Additional Implementation Details · p. 14

Tiny-VAE latent representation

  • The frozen Tiny-VAE encodes RGB images into 16-channel latents, and only the latent flow model is optimized.No additional latent normalization is applied; the same frozen decoder converts predicted latents back to images for evaluation.Source: C. Additional Implementation Details · p. 14

TVM warmup schedule

  • 5k/10k-step warmup and ramp for LSD-DLL, and 10k/20k for UHD-LOL4K, gradually introduce TVM before fixing its weight at 1.0.The TVM objective is disabled during warmup, introduced linearly during the ramp, and held at its final weight thereafter.Source: C. Additional Implementation Details · p. 14

D. Baseline Methods: Training Details

The baseline comparisons use official methods and training recipes where available, with dataset-specific retraining for LSD-DLL and UHD-LOL4K. The appendix also records individual baseline training protocols.

Baseline training protocol

  • LSD-DLL and UHD-LOL4K baselines were retrained using authors’ official code, architectures, optimization settings, and schedules, changing only dataset configuration as needed.For LOL-v1, official pretrained weights were used when available; TFFormer was trained from scratch because no official checkpoint was available.Source: D. Baseline Methods: Training Details · pp. 14–15

Baseline training details

  • 300,000 steps on 128×128 RGB patches train TFFormer, while HVI-CIDNet uses 1,000 epochs on 256×256 crops.TFFormer uses Adam with β1 =0.9 and β2 =0.99; HVI-CIDNet uses the default random-crop and flip augmentations.Source: D. Baseline Methods: Training Details · p. 15
  • 150K iterations on 256×256 patches train Multinex, while Retinexformer uses 2.5×105 iterations on 128×128 patches.Both use Adam and cosine annealing learning-rate schedules; their batch sizes are 8.Source: D. Baseline Methods: Training Details · p. 15
  • 500k iterations train PyDiff with its official 2,000-step diffusion schedule; WCDM follows its official schedule for 1,000 epochs.ReDDiT uses dataset-specific GSAD teachers followed by progressive distillation from eight-step to four-step and two-step students.Source: D. Baseline Methods: Training Details · p. 15

E. Additional Ablation Studies

The ablations examine inference budgets, one-step objectives, latent autoencoders, TVM warmup, and source-latent perturbation. Together, they characterize quality–compute trade-offs and factors affecting reconstruction and generalization.

Extended NFE ablation

  • With an 8.6× reduction in GFLOPs, one-step TVM retains slightly better PSNR and NIQE than No-TVM at 25 NFEs, while No-TVM achieves lower LPIPS.TVM-trained variants remain largely stable as NFEs increase, whereas No-TVM improves substantially from one to 25 NFEs.Source: E. Additional Ablation Studies · p. 16, Figure 6 · p. 16

One-Step Methods across NFEs

  • At one NFE, TVM provides the strongest joint distortion–perception performance; MeanFlow and ConsistencyFM improve NIQE with more NFEs but remain lower in PSNR.No-TVM requires many evaluations to approach one-step TVM, whose PSNR and NIQE remain stable across NFEs.Source: E. Additional Ablation Studies · p. 16, Figure 8 · p. 16

Autoencoder Reconstruction Analysis

  • 471 to 48 GFLOPs and 205M to 124M parameters accompany the TAESD-based variant’s higher PSNR and SSIM and lower NIQE than the SD-VAE-based variant.The SD-VAE variant has lower LPIPS and higher MUSIQ, reflecting a different distortion–perception trade-off.Source: E. Additional Ablation Studies · p. 16, Table 7 · p. 17
  • On LSD-DLL, direct TAESD reconstruction reaches 29.49dB PSNR and 0.069 LPIPS versus 20.3dB and 0.48 for GLOW-FM, indicating reconstruction headroom.Direct reconstruction isolates the autoencoder from the transport model and is a reference, not a strict bound on GLOW-FM output.Source: E. Additional Ablation Studies · pp. 16–17, Table 8 · p. 17

Autoencoder reconstruction error

  • 1.181, 1.792, and 0.266 ×10−3 MSE for TAESD versus 1.963, 4.579, and 0.693 for SD-VAE on LOL-v1, LSD-DLL, and UHD-LOL4K, respectively, measure only the autoencoder term, not the complete W2 bound.TAESD’s reconstruction term is smaller on all three benchmarks; the complete bound’s transport weighting and decoder-dependent constants are not estimated.Source: E. Additional Ablation Studies · p. 17, Table 9 · p. 17

Autoencoder reconstruction analysis

  • 16 latent channels per spatial location give TAESD four times as many latent coefficients as SD-VAE at the same downsampling factor.The passage connects this higher latent dimensionality with TAESD’s stronger reconstruction despite its smaller model size.Source: E. Additional Ablation Studies · p. 17

TVM Warmup Ablation

  • Gradually introducing TVM after flow matching establishes a reliable transport improves PSNR and yields substantially better NIQE at convergence than training without warmup.Without warmup, the displacement objective acts on an immature velocity field and leads optimization toward an inferior solution.Source: E. Additional Ablation Studies · p. 17, Figure 9 · p. 17

Effect of source-latent perturbation

  • σ=0.01 improves LSD-DLL PSNR and LPIPS over σ=0 and gives the best generalization on LSD-U and cross-dataset mean NIQE across five benchmarks.Although σ=0.10 has better in-domain NIQE, σ=0.01 gives the best overall balance across distortion, perceptual quality, and generalization.Source: E. Additional Ablation Studies · pp. 17–18, Table 10 · p. 18

F. Additional Quantitative Results

The extended comparisons report five-seed averages and broaden metric coverage. Runtime and memory measurements indicate efficiency across H100 GPUs and Arm SME/SME2 execution.

Additional quantitative results

  • Five evaluation seeds are summarized as mean±std, with CLIP-IQA added across benchmarks and SSIM and LPIPS added for LOL-v1.Extended results cover LOL-v1, LSD-DLL, UHD-LOL4K, and LSD-U.Source: F. Additional Quantitative Results · p. 18

Runtime and Memory Analysis

  • GLOW-FM-S has the lowest latency and peak allocated memory on H100 across evaluated resolutions, and all GLOW-FM variants have lower latency than competitors under Arm SME/SME2.Measurements use batch size one and average 100 timed inferences after 20 warm-up iterations; H100 tests span 256×256 to 4096×4096, with an additional 1024×1024 Apple M4 Max test.Source: F. Additional Quantitative Results · p. 18

G. Additional Qualitative Results

Additional qualitative comparisons show GLOW-FM’s outputs across benchmarks and resolutions, with natural illumination, preserved structure and detail, and no severe visual artifacts.

G. Additional Qualitative Results

Loading 2610.31222v1…