Geometry-Aware Image Flow Matching

A new preprint challenges the long-standing Euclidean assumptions in natural image generation, proposing that image semantics are better modeled on a hypersphere and introducing two geometry-aware flow matching methods that outperform standard baselines. The paper, submitted on 24 May 2026 to arXiv, argues that the field of natural image generation has remained confined to Euclidean assumptions, failing to exploit intrinsic geometric structures [1]. The authors investigate the geometry of natural images and report that semantic information is predominantly encoded in directional components, while norm components can be approximated by the global average [2]. This property holds across both RGB and latent spaces, suggesting that natural images can be effectively modeled on a hypersphere [2]. Building on this finding, the researchers introduce Spherical Optimal Transport Flow Matching (SOT-CFM), which utilizes angular distance, and Spherical Flow Matching (SFM), which constrains dynamics directly on the manifold [1]. Their experiments demonstrate that these geometry-aware methods achieve superior performance against Euclidean baselines [2]. The work aims to bridge the gap between Riemannian manifold-based modeling and natural image generation [1]. Flow matching is a technique related to the broader field of optical flow, which describes the pattern of apparent motion of objects, surfaces, and edges in a visual scene [3]. The concept has roots in Second World War research into pilot vision, with James J. Gibson publishing his theory in 1947 and coining the term "optic flow" in 1950 [3]. The term is also used in robotics for tasks including motion detection and object segmentation [3]. The research sits within the discipline of computer vision, which concerns the theory behind artificial systems that extract information from images [4]. Subdisciplines include scene reconstruction, object detection, video tracking, and 3D pose estimation [4]. Computer vision models are constructed with the aid of geometry, physics, statistics, and learning theory [4]. Advances in generative image modeling have downstream implications for applications such as facial recognition systems, which pinpoint and measure facial features from digital images or video frames [5]. Such systems are categorized as biometrics and are deployed in video surveillance, law enforcement, and automatic indexing of images [5]. Their use has raised controversy, with claims that the systems violate privacy, make incorrect identifications, and encourage racial profiling, leading to bans in several U.S. cities and Meta Platforms shutting down its Facebook facial recognition system in 2021 [5].