This repository provides the official implementation of Scalable Diffusion Models with Transformers (DiT). It replaces the traditional U-Net backbone in diffusion models with a transformer operating on latent patches. The work demonstrates that DiTs with higher computational complexity (measured in Gflops) consistently achieve better FID scores. Pre-trained DiT-XL/2 models are provided that achieve state-of-the-art results on class-conditional ImageNet generation at both 256×256 and 512×512 resolutions.