GLOW-FM: Efficient Generative Low-Light Image Enhancement with One-Step Latent Flow Matching
GLOW-FM attacks the high inference cost of diffusion-based low-light image enhancement (LLIE) by replacing iterative sampling with a single-step latent flow. The core idea is to run Terminal Velocity Matching (TVM) in the compact latent space of a frozen Tiny-VAE, directly predicting the displacement from a degraded low-light latent to its well-exposed counterpart in one function evaluation. If its quality claims hold up, the result would be a generative LLIE method practical enough for edge-device deployment.
GLOW-FM is a solid, practically motivated contribution that convincingly demonstrates one-step generative LLIE at low computational cost. The efficiency gains are real—GLOW-FM-S hits 1 NFE and 34 GFLOPs—and the ablations validate TVM over competing one-step objectives. Still, the paper’s headline claims rest partly on brittle no-reference metrics for cross-dataset generalization, distortion performance regresses on smaller datasets, and the theoretical Wasserstein bound in Appendix B relies on an unverified Lipschitz decoder constant that limits its usefulness.
The central engineering argument—that latent TVM plus a frozen Tiny-VAE yields strong one-step generative enhancement at low cost—is well supported. Table 5 shows that among matched TAESD one-step variants, TVM achieves the strongest overall distortion and perceptual performance, and the TAESD latent space is validated not only for efficiency but also for quality: the paper notes that 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 in-distribution results on LSD-DLL (Table 1) and the transparent reporting of model sizes and training details further strengthen the empirical case.
The chief concern is that the cross-dataset generalization claim hinges on no-reference metrics NIQE and BRISQUE, which are known to correlate poorly with human judgments and can be biased toward brighter, more contrasted outputs regardless of artifacts. The paper asserts that GLOW-FM achieves the best generative NIQE and BRISQUE results on nearly all evaluated benchmarks, yet it provides no user studies or downstream-task validation to corroborate this. Additionally, the single-step quality ceiling is evident: the extended ablation shows No-TVM at 25 NFEs achieves lower LPIPS, suggesting TVM sacrifices some fidelity for speed. Finally, the theoretical Wasserstein bound relies on a Lipschitz assumption on the decoder that is never verified numerically, rendering the bound non-actionable.
For the paired benchmarks, the comparisons appear fair: the authors state that for LSD-DLL and UHD-LOL4K, we trained the baselines on the respective training sets, whereas all other results use the authors’ official code and checkpoints. This protocol supports the finding that GLOW-FM achieves the strongest overall performance across several distortion and perceptual-quality metrics on LSD-DLL. However, the cross-dataset evaluation in Table 4 is less controlled because the paper does not explicitly state that every baseline was retrained on LOL-v2-Syn under identical conditions; if some generative methods used different source data or pretraining, the NIQE/BRISQUE rankings may reflect data exposure rather than model generalization. The qualitative figures are illustrative but inevitably cherry-picked.
Reproducibility is mixed. On the positive side, the paper releases code and provides extensive hyperparameters, HDiT configurations, and the TVM warmup schedule in the appendices. On the negative side, training demands approximately 100k optimization steps on 8 NVIDIA H100 GPUs with large crops, creating a high resource barrier for independent replication. Wall-clock runtime is not reported, so real-world latency advantages beyond GFLOP counts remain unclear. Inference also depends on an externally maintained TAESD checkpoint without guaranteed API stability.
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.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.