WaiT for the Signal: Simple Frequency-Aware Flow-Matching

2026-07-30

AI Summary

This work addresses the challenge in high-resolution image generation of jointly optimizing global structure, local details, and texture fidelity, compounded by existing flow-matching methods that treat all spatial frequencies uniformly despite high-frequency components degrading into noise earlier during the reverse process. To this end, the authors propose WaiT—a wavelet-aware image Transformer—that leverages invertible wavelet decomposition to decouple generation into low- and high-frequency pathways. A frequency-aware delay mechanism is introduced, allowing high-frequency components to engage in flow matching only after low-frequency structures have stabilized. This approach achieves the first temporally decoupled multi-scale generation framework and is accompanied by a novel tri-axis native-resolution evaluation protocol. Experiments demonstrate that WaiT attains Pareto-optimal performance on ImageNet 512×512 with a pixel-space FID of 1.43 (1.3 for the 2B variant) while reducing sampling compute by 50%, and sets a new state of the art on Kinetics-600 video generation with an FVD of 0.84.

Abstract

As image generation models scale to ever higher resolutions, global coherence, local detail, and texture fidelity become critical axes for generation quality. However, standard flow matching treats all spatial frequencies uniformly, ignoring the natural frequency hierarchy where high-frequency bands become indistinguishable from pure noise far earlier than coarse structures. We introduce WaiT, a Wavelet-aware image Transformer that decomposes generation into coarse and fine bands via lossless wavelets. True to its name, the high-frequency bands wait for the signal: staying pure noise until coarse structure has emerged, then joining the flow for joint refinement. Since standard FID discards fine-grained detail through aggressive downsampling, we introduce a more stringent three-axis evaluation protocol to assess quality at native resolution. On ImageNet 512x512, WaiT achieves a pixel-space FID of 1.43 and is Pareto-optimal across all three axes, reducing sampling compute by up to 50%. With our largest 2B model, we set a new state-of-the-art FID of 1.3 for pixel-space models on ImageNet 512 resolution. Our formulation outperforms even the strongest latent-space models on texture fidelity, and scales seamlessly to high-resolution OpenImages and to video generation, achieving a state-of-the-art FVD of 0.84 on Kinetics-600 with no algorithmic modifications.

Problem

• image generation
• spatial frequency
• flow matching
• texture fidelity
• high-resolution

Innovation

• Wavelet-aware generation
• Frequency-aware flow matching
• Coarse-to-fine refinement
• Native-resolution evaluation
• Pixel-space generative modeling