Dynamic Chunking Diffusion Transformer

Researchers introduce Dynamic Chunking Diffusion Transformer (DC-DiT), a new AI model that adaptively processes images by allocating more computational resources to detail-rich regions and fewer to uniform backgrounds. The system improves image generation quality while reducing computational costs by up to 16x compared to traditional diffusion transformers.

• →DC-DiT adaptively compresses images into variable-length token sequences, spending more compute on detailed regions and less on uniform backgrounds.
• →The model learns meaningful visual segmentations without explicit supervision and adapts compression across diffusion timesteps.
• →DC-DiT shows consistent improvements in FID and Inception Score over baseline models at both 4x and 16x compression rates.
• →The system can be efficiently retrofitted to existing pretrained DiT models with up to 8x fewer training steps required.
• →The technique has potential applications beyond images, extending to video and 3D generation tasks.