GLOW‑FM brightens low‑light photos in a single function call, collapsing an entire diffusion pipeline into one evaluation. The twist is that this speedup does not come at the expense of perceptual fidelity, contrary to the long‑standing trade‑off between quality and latency.
Before GLOW‑FM, most generative low‑light image enhancement (LLIE) systems relied on diffusion models that required dozens of iterative denoising steps. Those multi‑step samplers delivered impressive visual results but imposed heavy compute burdens, making real‑time deployment impractical for edge devices or interactive applications.
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 ["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." [1]].
The small variant runs at just 34 GFLOPs while delivering competitive image quality and cross‑dataset generalization, performing competitively with multi‑step baselines on standard benchmarks ["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." [1]].
On an NVIDIA H100 accelerator, GLOW‑FM‑S records the lowest latency and smallest peak memory footprint across all tested resolutions ["On the H100, GLOW-FM-S achieves the lowest latency and peak allocated memory across all evaluated resolutions." [1]].
The paper’s evaluation is limited to static image datasets; it does not explore video streams or extreme noise conditions, leaving open whether a single‑step flow match can preserve temporal consistency or handle severely corrupted inputs. This suggests that future work should test GLOW‑FM on continuous capture pipelines and investigate hybrid schemes that combine TVM with dedicated denoising stages.
If the reported 34 GFLOP budget holds in production, developers building real‑time photo‑enhancement features can replace costly diffusion samplers with a single‑call GLOW‑FM model and rebenchmark existing solutions against its latency and memory advantages.
