Aligning Latent Geometry for Spherical Flow Matching in Image Generation

Hugging Face Daily Papers Papers

Summary

This paper proposes aligning latent geometry for spherical flow matching, projecting latents onto a fixed-radius sphere and using spherical linear interpolation to improve image generation quality, consistently improving FID on class-conditional ImageNet.

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, we show through component-swap probes that decoded perceptual and semantic content is carried predominantly by direction, with radius contributing much less. We 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 at every timestep, and their velocity targets are purely angular by construction. Under matched training, the method consistently improves class-conditional ImageNet-256 FID across different image tokenizers, leaves the diffusion architecture unchanged, and requires no auxiliary encoder or representation-alignment objective.
Original Article
View Cached Full Text

Cached at: 05/15/26, 08:26 PM

Paper page - Aligning Latent Geometry for Spherical Flow Matching in Image Generation

Source: https://huggingface.co/papers/2605.15193

Abstract

Geodesic flow matching improves image generation by projecting latents onto fixed radius spheres and using spherical linear interpolation instead of linear paths, preserving semantic content through angular components.

Latent flow matchingfor image generation usually transportsGaussian noisetovariational autoencoderlatents along linear paths. Both endpoints, however, concentrate in thinspherical shells, and a Euclidean chord leaves those shells even when preprocessing aligns their radii. By decomposing each latent token intoradial and angular components, we show through component-swap probes that decoded perceptual and semantic content is carried predominantly by direction, with radius contributing much less. We therefore project data latents onto a fixed token radius, use the radial projection ofGaussian noiseas the spherical prior, finetune the decoder with the encoder frozen, and replace linear interpolation withspherical linear interpolation. The resultinggeodesic pathsstay on the sphere at every timestep, and their velocity targets are purely angular by construction. Under matched training, the method consistently improves class-conditional ImageNet-256 FID across differentimage tokenizers, leaves thediffusion architectureunchanged, and requires no auxiliary encoder or representation-alignment objective.

View arXiv pageView PDFProject pageAdd to collection

Get this paper in your agent:

hf papers read 2605\.15193

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2605.15193 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2605.15193 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2605.15193 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

Geometry-Aware Image Flow Matching

Hugging Face Daily Papers

This paper introduces geometry-aware flow matching for natural images by treating them as points on a hypersphere, proposing SOT-CFM and SFM methods that improve generative modeling by leveraging the spherical structure of image data.

Efficient Image Synthesis with Sphere Latent Encoder

Hugging Face Daily Papers

This paper proposes Sphere Latent Encoder, an efficient few-step image generation framework that performs denoising entirely in a spherical latent space, achieving high-quality 256×256 images with significantly reduced computational cost and improved FID scores on ImageNet-1K.

Energy-Guided Flow Matching

Hugging Face Daily Papers

Energy-Guided Flow Matching improves generative image quality by using a moving endpoint and adaptive scheduling, achieving state-of-the-art FID scores with reduced training cost.