Definition

Diffusion Transformer (DiT)

A Diffusion Transformer, or DiT, replaces the U-Net backbone of early diffusion models with a Transformer. The payoff is predictable scaling: more parameters and compute yield proportionally better images, the same law that drove GPT. DALL-E 3, Sora, Flux, and Veo build on it, and video models extend it with attention across time as well as space. Same architecture as text, different tokens.