Diffusion Transformer
An architecture that replaces the traditional U-Net denoising backbone of diffusion models with a Transformer. Noisy samples and noise timesteps are fed to the Transformer as a token sequence, and self-attention is used to model global dependencies. This design scales more predictably with model size and compute than convolutional alternatives.
Core Facts
Diffusion Transformer (DiT) is a class of generative backbone networks that introduces the Transformer architecture into diffusion models, cutting noisy images into patches and linearly embedding them into token sequences, and using multi-layer Transformer blocks to predict noise or velocity, replacing the U-Net in traditional diffusion models.
Compared to earlier U-Net-based diffusion models, the DiT architecture has stronger temporal consistency when processing long-sequence video frames.
Timeline (last 90 days)
DiT-type architectures tend to have higher parameter utilization efficiency than U-Net, which explains how the 500-million-parameter scale can still maintain considerable image quality.
MOSS-Image uses the DT (Diffusion Transformer) architecture.
The DiT (Diffusion Transformer) toolkit is based on the low-end Transformer and is a core tool with a very high usage rate in the industry.
The study adopts a latent diffusion framework, coupling a linear autoencoder with a Diffusion Transformer: the autoencoder compresses the actor network parameters into a compact latent space, and the Diffusion Transformer then models that latent distribution.
Diffusion Transformer (DiT) is a class of generative backbone networks that introduces the Transformer architecture into diffusion models, cutting noisy images into patches and linearly embedding them into token sequences, and using multi-layer Transformer blocks to predict noise or velocity, replacing the U-Net in traditional diffusion models.
OpenAI's Sora adopts the Diffusion Transformer (DiT) architecture, combining the attention mechanism of Transformers with the diffusion process, enabling it to understand spatiotemporal coherence.
Sora is based on the Diffusion Transformer (DiT) architecture, which was proposed by William Peebles and Saining Xie of UC Berkeley in 2023.
The latest generation of editing models adopts the DiT (Diffusion Transformer) architecture, which processes image tokens and text tokens in a unified manner, thereby achieving finer cross-modal alignment.
2024年以来随着DiT（Diffusion Transformer）架构在视频生成中的广泛应用，开源模型质量开始大幅追赶闭源产品
All Facts (14)
Diffusion Transformer (DiT) is a class of generative backbone networks that introduces the Transformer architecture into diffusion models, cutting noisy images into patches and linearly embedding them into token sequences, and using multi-layer Transformer blocks to predict noise or velocity, replacing the U-Net in traditional diffusion models.
80%VerifiedCompared to earlier U-Net-based diffusion models, the DiT architecture has stronger temporal consistency when processing long-sequence video frames.
65%UnverifiedInception AI的Mercury系列模型采用了扩散式Transformer（Diffusion Transformer）架构，可以并行生成多个token
85%Unverified早期视频生成模型基于U-Net架构，而扩散变换器将Transformer的注意力机制引入扩散过程
85%UnverifiedSeedance 2.0的技术路线基于扩散变换器（Diffusion Transformer，DiT）架构
85%UnverifiedOpenAI Sora基于Diffusion Transformer架构，在视觉质量上取得突破但仍频繁出现液体无故消失、刚体穿透等物理幻觉
80%UnverifiedDiT-type architectures tend to have higher parameter utilization efficiency than U-Net, which explains how the 500-million-parameter scale can still maintain considerable image quality.
50%UnverifiedMOSS-Image uses the DT (Diffusion Transformer) architecture.
50%UnverifiedThe DiT (Diffusion Transformer) toolkit is based on the low-end Transformer and is a core tool with a very high usage rate in the industry.
50%UnverifiedThe study adopts a latent diffusion framework, coupling a linear autoencoder with a Diffusion Transformer: the autoencoder compresses the actor network parameters into a compact latent space, and the Diffusion Transformer then models that latent distribution.
50%UnverifiedOpenAI's Sora adopts the Diffusion Transformer (DiT) architecture, combining the attention mechanism of Transformers with the diffusion process, enabling it to understand spatiotemporal coherence.
50%UnverifiedSora is based on the Diffusion Transformer (DiT) architecture, which was proposed by William Peebles and Saining Xie of UC Berkeley in 2023.
50%UnverifiedThe latest generation of editing models adopts the DiT (Diffusion Transformer) architecture, which processes image tokens and text tokens in a unified manner, thereby achieving finer cross-modal alignment.
50%Unverified2024年以来随着DiT（Diffusion Transformer）架构在视频生成中的广泛应用，开源模型质量开始大幅追赶闭源产品
50%