![[2602.16968] DDiT: Dynamic Patch Scheduling for Efficient Diffusion Transformers](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.16968] DDiT: Dynamic Patch Scheduling for Efficient Diffusion Transformers
Summary
The paper presents DDiT, a novel approach for dynamic patch scheduling in diffusion transformers, enhancing efficiency in image and video generation while maintaining quality.
Why It Matters
As diffusion transformers become increasingly prevalent in generative tasks, optimizing their computational efficiency is crucial. This research addresses the inefficiencies of fixed tokenization, proposing a dynamic method that adapts to content complexity, which could significantly impact the deployment of AI in real-time applications.
Key Takeaways
- DDiT introduces dynamic tokenization to improve efficiency in diffusion transformers.
- The method adapts patch sizes based on content complexity and denoising timesteps.
- It achieves significant speedups (up to 3.52x) without sacrificing generation quality.
- Early timesteps utilize coarser patches, while finer patches are used in later iterations for detail refinement.
- This approach enhances the practical application of diffusion models in real-time scenarios.
Computer Science > Computer Vision and Pattern Recognition arXiv:2602.16968 (cs) [Submitted on 19 Feb 2026] Title:DDiT: Dynamic Patch Scheduling for Efficient Diffusion Transformers Authors:Dahye Kim, Deepti Ghadiyaram, Raghudeep Gadde View a PDF of the paper titled DDiT: Dynamic Patch Scheduling for Efficient Diffusion Transformers, by Dahye Kim and 2 other authors View PDF HTML (experimental) Abstract:Diffusion Transformers (DiTs) have achieved state-of-the-art performance in image and video generation, but their success comes at the cost of heavy computation. This inefficiency is largely due to the fixed tokenization process, which uses constant-sized patches throughout the entire denoising phase, regardless of the content's complexity. We propose dynamic tokenization, an efficient test-time strategy that varies patch sizes based on content complexity and the denoising timestep. Our key insight is that early timesteps only require coarser patches to model global structure, while later iterations demand finer (smaller-sized) patches to refine local details. During inference, our method dynamically reallocates patch sizes across denoising steps for image and video generation and substantially reduces cost while preserving perceptual generation quality. Extensive experiments demonstrate the effectiveness of our approach: it achieves up to $3.52\times$ and $3.2\times$ speedup on this http URL and Wan $2.1$, respectively, without compromising the generation quality and promp...