Diffusion Transformer
A class of generative backbone networks that brings the Transformer architecture into diffusion models. Noisy images are split into patches and linearly embedded as token sequences, and multiple Transformer blocks predict noise or velocity, replacing the conventional U-Net.
Timeline (last 90 days)
SoftCache is completely training-free and adds no parameters specific to any backbone network, so it can be directly applied to different diffusion Transformer architectures without hyperparameter tuning for each model.
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.
The paper validates across multiple concepts and multiple text-to-image diffusion Transformers, finding that activation trajectories deviate significantly from straight lines.
Mainstream text-to-video models (including Seedance, Kling, Sora, etc.) are all based on the diffusion Transformer architecture; each generation is an independent random sampling, and two generations with the same prompt can produce significantly different results.
All Facts (4)
SoftCache is completely training-free and adds no parameters specific to any backbone network, so it can be directly applied to different diffusion Transformer architectures without hyperparameter tuning for each model.
50%UnverifiedDiffusion 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.
50%UnverifiedThe paper validates across multiple concepts and multiple text-to-image diffusion Transformers, finding that activation trajectories deviate significantly from straight lines.
50%UnverifiedMainstream text-to-video models (including Seedance, Kling, Sora, etc.) are all based on the diffusion Transformer architecture; each generation is an independent random sampling, and two generations with the same prompt can produce significantly different results.
50%