WaiT image model hits 1.43 FID with 50% less sampling compute

A collaboration between Meta AI, ENS, and Collège de France researchers has introduced WaiT (Wavelet-aware image Transformer), a new flow-matching architecture that exploits the natural frequency hierarchy of images. Standard flow matching treats all spatial frequencies uniformly, which becomes inefficient at high resolutions because fine details are indistinguishable from pure noise much earlier than coarse structures. WaiT instead decomposes generation via lossless wavelets into coarse and fine bands — the high-frequency bands literally 'wait for the signal,' remaining pure noise until the coarse structure has formed, then joining the flow for joint refinement.

On ImageNet 512x512, WaiT achieves a pixel-space FID of 1.43, setting a new state-of-the-art for pixel-space models at this resolution, and its largest 2B parameter variant pushes that to 1.3 FID. The model is Pareto-optimal across three evaluation axes (global coherence, local detail, texture fidelity), reducing sampling compute by up to 50% compared to baselines. Notably, it also outperforms strong latent-space models on texture fidelity, and scales seamlessly to video generation — hitting a state-of-the-art FVD of 0.84 on Kinetics-600 with zero algorithmic modifications. The authors also criticize standard FID for discarding fine-grained detail via aggressive downsampling, proposing a three-axis protocol that judges quality at native resolution.