Pixel Space End-to-End Learning Hierarchy

An animation introduces HiT: image generation in pixel space, with end-to-end learning and a learned visual hierarchy.

Hierarchical Diffusion Transformer for End-to-End Image Generation

Xiaoyu Wu△,*,† Yifei Wang□,* Weihang Guo□,* Chen Wei□ Lydia E Kavraki□ Niloofar Mireshghallah△ Zhiwei Steven Wu△
△ Carnegie Mellon University □ Rice University

*Equal contribution. †Project lead.

ImageNet

Better image quality.
Fewer FLOPs.

HiT reaches the same image quality with fewer training FLOPs.

ImageNet · Generated samples

HiT learns its own
visual hierarchy.

HiT generates images in pixel space. Patch groups adapt to each image and diffusion timestep.

Text to image · Generated samples

Images and words.
One hierarchy.

See patch groups and word connections in images generated from text prompts.

Input images · Word connections

Just for fun.

HiT connects image regions and caption words through a shared hierarchy.

These connections emerge without word–region annotations.

BibTeX

@article{wu2026hierarchical,
  title={Hierarchical Diffusion Transformer for End-to-End Image Generation},
  author={Wu, Xiaoyu and Wang, Yifei and Guo, Weihang and Wei, Chen and Kavraki, Lydia E and Mireshghallah, Niloofar and Wu, Steven},
  journal={arXiv preprint arXiv:2611.22962},
  year={2026}
}