Reproduce “FlexiDiT: Your Diffusion Transformer Can Easily Generate High-Quality Samples with Less Compute” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Sotiris Anagnostidis, Gregor Bachmann, Yeongmin Kim, Jonas Kohler, Markos Georgopoulos, Artsiom Sanakoyeu, Yuming Du, Albert Pumarola and 2 more

Meta GenAI and ETH Zürich researchers introduce FlexiDiT, a breakthrough framework that enables Diffusion Transformers to generate high-quality images with up to 40% less compute through dynamic resource allocation and flexible tokenization, while maintaining state-of-the-art generation quality and extending naturally to video generation with 75% compute savings.

Key insights

  • Different stages of the diffusion process require varying levels of computational resources
  • A single flexible model can effectively handle multiple compute configurations without quality loss
  • Dynamic resource allocation principles extend naturally across different modalities and architectures

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.