Scalable Diffusion Models with Transformers (DiT)

Replaced standard convolutional U-Net backbones in latent diffusion models with transformer blocks conditioning on time and class via adaptive layer norm (adaLN-Zero).

Directly guides architectural decisions, alignment strategy, and serving infrastructure for task-image-generation, task-video-generation.

Assumptions

• Empirical distribution regularity holds and target domain adheres to pretraining linguistic/visual support

Limitations

• Resource scaling, inference memory requirements, and alignment robustness vary with model size and hardware topology