Underwater Image Enhancement by Transformer-based Diffusion Model with Non-uniform Sampling for Skip Strategy

2 min read

This paper enhances underwater images with a conditional denoising diffusion model: the degraded image and Gaussian noise go in, and the network denoises towards a clean image. To make the reverse process cheaper it uses a lightweight transformer denoiser, so each step is faster, and a skip-sampling strategy, so there are fewer steps. The skipped time steps are chosen non-uniformly, either piecewise or by an evolutionary search over step sequences.

Findings

  • The abstract reports performance competitive with recent methods on the usual paired enhancement datasets, at higher efficiency; it gives no headline number.
  • Non-uniform step schedules beat uniform skipping at the same number of steps.

Limits

  • I read the abstract and the repository README only.
  • Trained on paired data, so it learns what the reference images in those datasets look like, which is not the same as removing the water physically.
  • Built for appearance. Nothing in what I read tests whether enhanced frames give better matches or depth.

Relevance

DM_underwater is the enhancer inside MERLION-E, where it runs on 128x128 crops to fit a Jetson. Whether learned enhancement helps geometry or only looks better is still open; compare summers-2025-enhancement-matching and alvareztunon-2026-gs-underwater. See underwater perception.