DiT
A diffusion model architecture proposed by Meta AI in 2023 that replaces the traditional U-Net backbone with a Vision Transformer. It exhibits strong scalability with large parameter counts and has been adopted by advanced video generation models such as Sora.
Core Facts
DiT was proposed by Peebles and Xie in 2022, introducing the Transformer architecture into the diffusion process.
H3 Minimax is Minimax's video generation engine, which adopts a Sparse Attention mechanism.
Sora adopts the DiT (Diffusion Transformer) architecture, compressing video into spatiotemporal patch sequences before feeding them into a Transformer, replacing the traditional U-Net with a Transformer as the denoising backbone network.
Timeline (last 90 days)
InstantFusion's shared interface makes cross-architecture online policy distillation possible, allowing a teacher model (e.g., a large DiT) and a student model (e.g., a lightweight UNet) to exchange intermediate states in a shared latent space.
DiT replaces convolutional layers with Transformer blocks, outperforming earlier diffusion models that used UNet architectures in high-resolution generation.
If one wishes for a DiT to serve both as a generative model and as a high-quality feature extractor, introducing long skip connections is a principled design choice.
Although DiT achieves higher generation quality than U-Net, the discriminative features it extracts are actually weaker.
DiT with long skip connections (LSC) can produce higher-quality features.
Through controlled training-from-scratch experiments, the study isolated long skip connections (LSC) as the core factor influencing the quality of DiT features.
Stable Diffusion 3.5 introduces a Transformer backbone (DiT architecture), with improvements in both parameter scale and generation quality compared to its predecessor.
The core of Nanosaur 2 is a 670M-parameter DiT (Diffusion Transformer) architecture, paired with a brand-new VAE and a 270M-parameter text encoder.
The SupraLabs team has released a lightweight text-to-image model, Supra2-IMG, with 100M parameters, trained from scratch using a DiT architecture, with an output resolution of 256×256, and it is fully open-source.
The weights of image generation models are typically divided into multiple components such as a text encoder, a diffusion backbone network (UNet or DiT), and a decoder (VAE).
26 more timeline events
All Facts (20)
DiT was proposed by Peebles and Xie in 2022, introducing the Transformer architecture into the diffusion process.
90%VerifiedH3 Minimax is Minimax's video generation engine, which adopts a Sparse Attention mechanism.
80%VerifiedSora adopts the DiT (Diffusion Transformer) architecture, compressing video into spatiotemporal patch sequences before feeding them into a Transformer, replacing the traditional U-Net with a Transformer as the denoising backbone network.
80%VerifiedThe Wan (万象) video generation model is based on the DiT (Diffusion Transformer) architecture.
80%VerifiedFLUX adopts a DiT (Diffusion Transformer) architecture, replacing the traditional U-Net structure with a pure Transformer.
80%VerifiedDiffusion Transformer (DiT) is the dominant architecture in the current image generation field, used in models such as Stable Diffusion 3 and FLUX.
75%VerifiedThe denoising process of diffusion models typically requires tens to hundreds of iterations (called sampling steps), each of which requires a complete forward pass through the neural network.
75%UnverifiedTransformer架构被引入视觉生成（如Vision Transformer、DiT架构），使模型能更好理解图像不同区域间的全局关系，适用于多模态场景中文本token和图像token的统一处理
88%UnverifiedDiT is the foundational architecture of next-generation generative models such as Stable Diffusion 3 and FLUX.
70%UnverifiedOpenAI's Sora is widely believed to have adopted an architecture similar to DiT.
60%UnverifiedMainstream video generation models such as Sora and CogVideoX widely adopt the DiT or its variant architectures, marking the transition of video generation technology from the UNet era to the Transformer era.
60%UnverifiedThe DiT architecture has become the common technical choice for mainstream video generation models such as Sora, Kling, and Wan.
60%UnverifiedThe DiT architecture replaces the traditional U-Net with a Transformer as the diffusion model backbone, enabling the capture of long-range cross-frame dependencies at the patch level.
60%UnverifiedOpenAI's Sora and Google's VideoPoet both adopt the technical approach of compressing video frames into discrete tokens followed by autoregressive prediction
60%UnverifiedVideo generation models such as OpenAI's Sora and Kuaishou's Kling also adopt a technical approach similar to DiT.
60%UnverifiedDiT replaces convolutional layers with Transformer blocks, outperforming earlier diffusion models that used UNet architectures in high-resolution generation.
50%UnverifiedInstantFusion's shared interface makes cross-architecture online policy distillation possible, allowing a teacher model (e.g., a large DiT) and a student model (e.g., a lightweight UNet) to exchange intermediate states in a shared latent space.
50%UnverifiedAlthough DiT achieves higher generation quality than U-Net, the discriminative features it extracts are actually weaker.
50%UnverifiedDiT with long skip connections (LSC) can produce higher-quality features.
50%UnverifiedThrough controlled training-from-scratch experiments, the study isolated long skip connections (LSC) as the core factor influencing the quality of DiT features.
50%