Reproduce “CurveFlow: Curvature-Guided Flow Matching for Image Generation” 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

Yan Luo, Drake Du, Hao Huang, Yi Fang, Mengyu Wang

CurveFlow introduces a curvature-guided flow matching framework that learns smooth, non-linear trajectories for image generation, overcoming the limitations of linear paths in existing rectified flow models. The method achieves superior semantic consistency and instructional compliance, alongside state-of-the-art image quality as indicated by the lowest FID scores on MS COCO datasets.

Key insights

  • Strictly linear generative trajectories are suboptimal for accurately modeling complex, non-linear data manifolds and preserving fine-grained semantic information in generated content.
  • Explicitly modeling and regularizing the curvature of flow matching trajectories significantly enhances both the semantic fidelity (instructional compliance) and visual quality of generated images.
  • The geometric path taken by a generative model through the data manifold is a critical factor influencing the meaningfulness and fidelity of its output, rather than merely a computational detail.

Summary by alphaXiv.

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