Research

Hierarchical Diffusion Transformer (HiT) Tackles End-to-End Image Generation in Pixel Space

2 days ago
HiT: a beaver holding a mug, beside its saved Level 2 hard groups at diffusion time 0.85.

Image via nicholas0228.github.io

Researchers from Carnegie Mellon University and Rice University have introduced HiT, a hierarchical diffusion transformer designed for end-to-end image generation directly in pixel space. The model learns visual hierarchies from scratch rather than relying on latent compression stages common in many existing diffusion architectures.

The work, led by Xiaoyu Wu and co-authored by Yifei Wang, Weihang Guo, Chen Wei, Lydia E. Kavraki, Niloofar Mireshghallah, and Zhiwei Steven Wu, is presented as an arXiv preprint. By operating end-to-end in pixel space with learned visual hierarchies, HiT aims to offer a more unified and interpretable approach to image synthesis compared to pipeline-based alternatives.