WaiT: Wavelet-Aware Image Transformer Achieves Pareto-Optimal FID
A new study on arXiv (2607.28760) introduces WaiT, a unique image Transformer that uses wavelets to distinguish between coarse and fine details in image generation. Unlike traditional flow matching methods that treat all frequencies the same, WaiT keeps high-frequency details as noise until the broader structure is established, then refines them together. The researchers also created a robust evaluation method with three axes to assess image quality at the original resolution, overcoming the shortcomings of typical FID metrics that often miss finer details due to downsampling. On ImageNet 512x512, WaiT achieved a pixel-space FID of 1.43 and proved Pareto-optimal across all metrics. The paper is classified as cross-announcement and is available online.
Key facts
- WaiT is a Wavelet-aware image Transformer for image generation.
- It decomposes generation into coarse and fine bands via lossless wavelets.
- High-frequency bands wait for the signal: staying pure noise until coarse structure emerges.
- Standard flow matching treats all spatial frequencies uniformly.
- A new three-axis evaluation protocol is introduced to assess quality at native resolution.
- On ImageNet 512x512, WaiT achieves a pixel-space FID of 1.43.
- WaiT is Pareto-optimal across all thresholds.
- The paper is available on arXiv with ID 2607.28760.
Entities
—