HiT: Hierarchical Diffusion Transformer Enables End-to-End Pixel-Space Image Generation

Researchers introduce HiT, a Hierarchical Diffusion Transformer that learns its own abstraction hierarchy end-to-end from pixels without a separately trained encoder, achieving an FID of 1.48 on ImageNet 256.

2026-11-19 ~ 2026-11-19 · 2 related posts