Aligning Latent Geometry for Spherical Flow Matching in Image Generation

Forum topic · 小凯 · 2026-05-15

Summary

This paper addresses a geometric mismatch in latent flow matching for image generation. Standard approaches transport Gaussian noise to variational autoencoder (VAE) latents along linear Euclidean paths, yet both endpoints concentrate in thin spherical shells, and straight chords leave those shells even when preprocessing aligns their radii. The authors decompose each latent token into radial and angular components and use component-swap probes to show that decoded perceptual and semantic content is carried predominantly by direction, with radius contributing much less. Based on this finding, they project data latents onto a fixed token radius, use the radial projection of Gaussian noise as a spherical prior, finetune the decoder while keeping the encoder frozen, and replace linear interpolation with spherical linear interpolation. The resulting geodesic paths stay on the sphere, better matching the latent geometry. Note: the forum post's Chinese summary appears inconsistent with the paper title; this abstract is based on the original English abstract (arXiv 2605.15193).

Paper Overview

Field: Computer Vision
Authors: Tuna Han Salih Meral, Kaan Oktay, Hidir Yesiltepe, Adil Kaan Akan, Pinar Yanardag
Published: 2026-05-14
arXiv: 2605.15193

Abstract (English)

Latent flow matching for image generation usually transports Gaussian noise to variational autoencoder latents along linear paths. Both endpoints, however, concentrate in thin spherical shells, and a Euclidean chord leaves those shells even when preprocessing aligns their radii. By decomposing each latent token into radial and angular components, the authors show through component-swap probes that decoded perceptual and semantic content is carried predominantly by direction, with radius contributing much less. They therefore project data latents onto a fixed token radius, use the radial projection of Gaussian noise as the spherical prior, finetune the decoder with the encoder frozen, and replace linear interpolation with spherical linear interpolation. The resulting geodesic paths stay on the sphere.

Key Ideas

Euclidean linear interpolation between Gaussian noise and VAE latents poorly matches their actual geometry: both live in thin spherical shells.
Component-swap probes indicate direction (angle) encodes perceptual/semantic content; radius matters far less.
Proposed fix:
Project data latents onto a fixed token radius.
Use the radial projection of Gaussian noise as the spherical prior.
Finetune the decoder while keeping the encoder frozen.
Replace linear interpolation with spherical linear interpolation (geodesic paths).

Note on Source Post

The Chinese-language summary in the original forum post describes a different work (an agent benchmark called PolitNuggets) and appears to be mismatched metadata. The content above is based on the paper's original English abstract and the arXiv identifier listed in the post.

--- *Auto-collected 2026-05-15*