Glossary · Term

Diffusion Transformer

← all terms

Definition

Plain language

The standard neural network design behind modern image and video generators.

As stated in the literature

A transformer (DiT) that operates over continuous latent patches and is trained with a diffusion or flow-matching objective; the backbone of Stable Diffusion 3 and large video models, transplanted to text in TextLDM.

Also called: DiT

Why it matters: It is the shared backbone behind today's leading image and video generators, so adapting it to other media lets those systems reuse a proven design.

For example, this is the kind of network that takes a text prompt and a field of random noise and turns it into a detailed picture.

Heard on the show

“… with a variational autoencoder, you run a plain transformer over those latents — that's the Diffusion Transformer, the DiT — you train it with something called flow matching, and you steer generation with …”
Episode 127 — What Diffusion Language Models Were Missing: A Map, Not an Algorithm

Mentioned in 1 episode

  1. 127
    What Diffusion Language Models Were Missing: A Map, Not an Algorithm

Related terms