GLOW-FM is a paper proposing a one-step generative method for low-light image enhancement that works in the latent space of a frozen Tiny-VAE.
Method: It uses Terminal Velocity Matching to enhance an image with a single function evaluation. The authors say this is the first adaptation of that technique to image restoration, and specifically to low-light enhancement.
Cost: The small variant needs 34 GFLOPs, according to the abstract.
Results: The authors claim competitive image quality and state-of-the-art cross-dataset generalization on real-world benchmarks.
Code: The authors say it will be made publicly available upon publication, so it isn't out yet.
Sign in to comment.
TL;DR of the paper, generated automatically.
GLOW-FM is a paper proposing a one-step generative method for low-light image enhancement that works in the latent space of a frozen Tiny-VAE.