Scalable Diffusion Models with Transformers

TL;DR

This paper introduces Diffusion Transformers, a new class of scalable diffusion models using transformer architecture for image generation, achieving state-of-the-art results on ImageNet benchmarks.

Contribution

It replaces U-Net with transformers in latent diffusion models, analyzing scalability and demonstrating superior performance on image synthesis benchmarks.

Findings

Higher Gflops in DiTs lead to lower FID scores.

Largest DiT-XL/2 models outperform previous diffusion models.

Achieved state-of-the-art FID of 2.27 on ImageNet 256x256.

Abstract

We explore a new class of diffusion models based on the transformer architecture. We train latent diffusion models of images, replacing the commonly-used U-Net backbone with a transformer that operates on latent patches. We analyze the scalability of our Diffusion Transformers (DiTs) through the lens of forward pass complexity as measured by Gflops. We find that DiTs with higher Gflops -- through increased transformer depth/width or increased number of input tokens -- consistently have lower FID. In addition to possessing good scalability properties, our largest DiT-XL/2 models outperform all prior diffusion models on the class-conditional ImageNet 512x512 and 256x256 benchmarks, achieving a state-of-the-art FID of 2.27 on the latter.