Scalable Diffusion Models with Transformers (DiT)

Replaced traditional U-Net backbones in latent diffusion with Vision Transformers (DiT), proving predictable compute scaling laws for generative diffusion.

The core architectural backbone powering next-generation diffusion systems: OpenAI Sora, Stable Diffusion 3, Flux.1, and PixArt.

Assumptions

• Operating on sequences of latent patches with adaptive layer norm conditioning scales smoothly with increased model capacity and compute

Limitations

• Quadratic sequence scaling on high-resolution latent patch sequences without windowed or linear attention mechanisms