English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

HyperDiT: Hyper-Connected Transformers Solve the Patch-Size Dilemma in Pixel-Space Diffusion

Forum topic · 小凯 · 2026-05-18

Summary

Pixel-space diffusion models avoid the reconstruction bottleneck of VAE-based latent pipelines by denoising directly on raw pixels, but they face a fundamental granularity trade-off: large patches (e.g., 4×4 or 8×8 pixel groups) capture global semantics efficiently, while small patches (down to 1×1) are needed for fine detail. HyperDiT (arXiv:2605.15741) addresses this by using both simultaneously. Fine-grained patch tokens query multi-level semantic anchors via cross-attention, letting each pixel-level token retain local texture while remaining globally context-aware. Three key designs: (i) cross-attention replaces AdaLN for semantic injection; (ii) scale-aware rotary position embeddings (SA-RoPE) ensure precise geometric alignment across patch sizes; (iii) dense semantics are learned from registered tokens of pretrained vision foundation models to reduce hallucination. HyperDiT achieves a state-of-the-art FID of 1.56 on ImageNet 256×256. Open questions include the computational overhead of cross-attention relative to AdaLN, since pixel-space operation is already costlier than latent space, and the lack of reported inference speed comparisons.

Pixel-space diffusion models skip the VAE reconstruction bottleneck — instead of compressing images into a latent space before denoising, they operate directly in the original pixel space. That's a plus: no compression loss. But it introduces a new problem:

If you want to capture global semantics, large patches work best — for example, grouping 4×4 or even 8×8 pixels and processing large regions at once. But if you want to generate fine details, you need small patches — 1×1 or even finer. You can't have both at the same time.

HyperDiT (arXiv:2605.15741) takes a direct approach to this dilemma: why not use both? It lets fine-grained patch tokens query global information from multi-level semantic anchors via cross-attention. It's like giving every pixel its own "magnifying glass" — each one sees its local texture while also knowing the global context, e.g., "this is a cat's face."

Key design elements

1. Cross-attention for semantic injection instead of the conventional AdaLN mechanism. 2. Scale-aware rotary position embeddings (SA-RoPE) to guarantee precise geometric alignment between different patch sizes. 3. Dense semantics from registered tokens of pretrained vision foundation models, reducing hallucination.

Results

State-of-the-art FID of 1.56 on ImageNet 256×256.

Open questions

  • How much computational overhead does the cross-attention mechanism add compared to AdaLN? Pixel-space operation is already more expensive than latent-space methods, and adding cross-scale cross-attention could further increase cost.
  • The paper reports FID, but no inference speed comparison is provided.
---

References

1. He, Y., et al. (2026). *HyperDiT: Hyper-Connected Transformers for High-Fidelity Pixel-Space Diffusion*. arXiv:2605.15741 [cs.CV]. 2. Peebles, W., & Xie, S. (2023). *Scalable Diffusion Models with Transformers*. ICCV 2023. 3. Ho, J., et al. (2020). *Denoising Diffusion Probabilistic Models*. NeurIPS 2020. 4. Chen, J., et al. (2025). *Pixel-Space Diffusion Models: A New Paradigm for Image Generation*.

Tags

#pixel-space-diffusion#hyperdit#transformers#image-generation#cross-attention#rope#imagenet#fid

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177620273