Browse Teach Curate

Diffusion Transformer Architecture in Deep Learning

The Diffusion Transformer (DiT) replaces the U-Net backbone of a diffusion model with a transformer operating on sequences of latent patches, conditioning on the diffusion timestep and on text or class inputs through adaptive normalization. Students learn how patchifying latents into tokens, self-attention over those tokens, and conditioning mechanisms let diffusion models scale like transformers.

jeremy Video

Diffusion Transformers in Vision and Image Generation

The core principle involves adapting transformer architectures for vision tasks by transitioning from unidirectional causal attention to bidirectional processing and reducing computational complexity…