Geometry-Aware Image Flow Matching

Researchers propose spherical flow-matching methods for image generation, modeling natural images on a hypersphere instead of in Euclidean space.

The paper argues that much of an image’s semantic information is captured by directional components, making angular geometry especially relevant.

Two methods are introduced: Spherical Optimal Transport Flow Matching (SOT-CFM) and Spherical Flow Matching (SFM).

Both use manifold-constrained or angular dynamics and reportedly outperform standard Euclidean baselines.

The work suggests intrinsic geometric structure could improve generative quality and broaden future image modeling approaches.