WaiT: Wavelet-Aware Flow-Matching Sets New SOTA, Halves Compute

CSProfKGD · x · 2026-08-04

The paper WaiT for the Signal introduces a Wavelet-aware image Transformer (WaiT). While standard flow matching treats all spatial frequencies uniformly, WaiT decomposes generation into coarse and fine bands via lossless wavelets.

Its core mechanism makes high-frequency bands 'wait' for the signal: staying as pure noise until coarse structures emerge, then joining the flow for joint refinement. Experiments show that on ImageNet 512x512, this method achieves a pixel-space FID of 1.43, is Pareto-optimal, and reduces sampling compute by up to 50%. With a 2B parameter model, it sets a new SOTA FID of 1.3. The architecture also scales seamlessly to high-res OpenImages and video generation, achieving an SOTA FVD of 0.84 on Kinetics-600.

Related event: WaiT Introduces Wavelet-Aware Flow Matching, Halving Compute and Setting FID Record(3 posts)→

Original post →

More from Multimodal

Multimodal channel →