Reproduce “Towards Hierarchical Rectified Flow” 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

Yichi Zhang, Yici Yan, Alex Schwing, Zhizhen Zhao

Researchers at the University of Illinois Urbana-Champaign developed Hierarchical Rectified Flow (HRF), a generative modeling framework that models multimodal velocity distributions using coupled ODEs, resulting in straighter integration paths. This method improves sampling efficiency by reducing Neural Function Evaluations (NFEs) and enhances sample quality, demonstrating better FID scores on datasets such as MNIST and CIFAR-10 compared to existing Rectified Flow baselines.

Key insights

  • The true 'ground-truth' velocity distribution for generative processes can be highly multimodal, and capturing this is crucial for more accurate data representation.
  • By modeling the full multimodal velocity distribution, HRF enables integration paths to intersect, leading to significantly straighter generative trajectories.
  • The sampling process for HRF is formally characterized as a Random Differential Equation (RDE) with a learned random velocity field, providing a novel theoretical perspective for ODE-based generative models.

Summary by alphaXiv.

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