$\Delta$-DiT: Accelerating Diffusion Transformers without training via Denoising Property Alignment
Keywords: Diffusion Model, Training-Free, Acceleration, Diffusion Transformer
Abstract: Diffusion models are now commonly used for producing high-quality and diverse images, but the iterative denoising process is time-intensive, limiting their usage in real-time applications. As a result, various acceleration techniques have been developed, though these primarily target UNet-based architectures and are not directly applicable to Transformer-based diffusion models (DiT). To address the specific challenges of the DiT architecture, we first analyze the relationship between the depth of DiT blocks and the quality of image generation. While skipping blocks can lead to large degradations in generation quality, we propose the $\Delta$-Cache method, which captures and stores the incremental changes of different blocks, thereby mitigating the performance gap and maintaining closer alignment with the original results. Our analysis indicates that the shallow DiT blocks primarily define the global structure of images such as compositions, and outlines, while the deep blocks refine details. Based on this, we introduce a denoising property alignment method that selectively bypasses computations of different blocks at various timesteps while preserving performance. Comprehensive experiments on PIXART-$\alpha$ and DiT-XL demonstrate that $\Delta$-DiT achieves a $1.6\times$ speedup in 20-step generation and enhances performance in most cases. In the 4-step consistent model generation scenario, and with a more demanding $1.12\times$ acceleration, our approach significantly outperforms existing methods.
Primary Area: generative models
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 9904
Loading