Introduction
As image generation models scale to increasingly higher resolutions, three factors become critical for output quality: global coherence, local detail, and texture fidelity. Yet standard flow-matching approaches treat all spatial frequencies uniformly, overlooking the natural frequency hierarchy in images. In particular, high-frequency bands become indistinguishable from pure noise far earlier than coarse structures, yet are modeled with the same intensity throughout the generation process.
In response, we introduce WaiT, a Wavelet-aware image Transformer that decomposes generation into coarse and fine bands using lossless wavelets. True to its name, the high-frequency bands “wait for the signal”: they remain pure noise until coarse structures have emerged, after which they join the flow for joint refinement. This staged approach aligns with the information content of each frequency band, improving both efficiency and quality.
Given that standard FID metrics discard fine-grained details through aggressive downsampling, we also introduce a more stringent three-axis evaluation protocol to assess quality at native resolution. On ImageNet 512×512, 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 at this resolution. Our formulation outperforms even the strongest latent-space models on texture fidelity and scales seamlessly to high-resolution OpenImages and video generation, achieving a state-of-the-art FVD of 0.84 on Kinetics-600 with no algorithmic modifications.
Methods
Frequency-Aware Decomposition
We employ a lossless wavelet transform to separate input data into coarse (low-frequency) and fine (high-frequency) bands. This decomposition respects the natural hierarchy of image information: low-frequency components capture global structure and lighting, while high-frequency components contain edges, textures, and fine details. Unlike prior methods that treat all bands uniformly, WaiT models each band according to its information density, which changes over the course of generation.
Cascade Sampling with Waiting
Our core innovation is a cascade sampling scheme where high-frequency bands "wait" until coarse structures are sufficiently formed. Initially, only the low-frequency band is active, allowing the model to establish global coherence without being distracted by noise-like high frequencies. Once the coarse flow has advanced enough, the fine bands join, enabling joint refinement. This waiting mechanism reduces redundant computation and improves training stability, as the model learns to condition fine details on established structures.
Three-Axis Evaluation Protocol
Standard FID computes features after aggressive downsampling (e.g., to 299×299), which discards high-frequency detail. To better capture true generation quality at native resolution, we propose a three-axis evaluation protocol that measures: (1) global coherence, (2) local structure fidelity, and (3) texture realism. Each axis is evaluated using dedicated metrics that are sensitive to frequency-specific errors, providing a more nuanced picture of model performance.
Results
Image Generation
On the ImageNet 512×512 benchmark, WaiT achieves a pixel-space FID of 1.43, surpassing previous pixel-space models. Our largest model (2B parameters) reaches an FID of 1.3, setting a new state-of-the-art among pixel-space approaches. Notably, WaiT is Pareto-optimal across all three evaluation axes, meaning no other model simultaneously achieves better coherence, local fidelity, and texture realism. Furthermore, the waiting mechanism reduces sampling compute by up to 50% relative to standard flow matching, without any loss in quality.
Texture Fidelity
On texture-centric metrics, WaiT outperforms even the strongest latent-space models. This advantage arises from our explicit modeling of high-frequency bands, which are often degraded by latent-space compression. By preserving these details in the pixel domain, we achieve sharper edges and more realistic surface patterns.
Scalability and Video Generation
Beyond static images, WaiT scales seamlessly to high-resolution videos. Without any algorithmic changes, we apply the same waiting scheme to spatiotemporal wavelets, achieving a state-of-the-art FVD of 0.84 on Kinetics-600. This shows that frequency-aware waiting is a general principle applicable across modalities.
Conclusion
We have presented WaiT, a frequency-aware flow-matching framework that leverages the natural hierarchy of image frequencies. By letting high-frequency bands wait for coarse structures to emerge, we improve generation quality and reduce computational cost simultaneously. Our results on ImageNet and Kinetics-600 demonstrate state-of-the-art performance across multiple benchmarks, and our three-axis evaluation protocol offers a more comprehensive measure of generation fidelity. Future work will explore adaptive waiting schedules that depend on input complexity and further scaling to even higher resolutions.
via ArXiv CV