Summary
This arXiv paper (2611.09863) introduces FlowAlign, a flow-guided method for multimodal alignment that maps a source modality into a fixed target-modality space. The key insight is that existing objectives address only part of the alignment goal: contrastive losses encourage relative similarity, distribution-matching losses improve marginal agreement, and pointwise regression fits source samples to paired targets without modeling target-space path geometry. FlowAlign repurposes conditional flow matching as a training-time alignment signal rather than an inference-time transport model, using a learned flow field to guide source representations so that they preserve paired correspondence while remaining compatible with the global structure of the target space. The authors provide theoretical analysis showing how FlowAlign improves paired correspondence by controlling endpoint accuracy and capturing local target-space geometry. Experiments on synthetic data, paired single-cell multi-omics benchmarks, and image-text alignment demonstrate that FlowAlign improves instance-level correspondence across diverse modality settings while preserving target-space structure. The work is relevant to computer vision, representation learning, and cross-modal alignment research.
Paper Overview
- Field: Computer Vision
- Authors: Ruizhi Yuan, Zeqiu Yu, Wei Gao, Wei Chen, Xiwei Tang
- Published: 2026-11-06
- arXiv: 2611.09863
Summary
Many multimodal alignment problems require mapping a source modality into a fixed space defined by a target modality, where the target space is not merely an intermediate embedding but is used directly for downstream analysis. In this setting, successful alignment should preserve both the global structure of the target distribution and the instance-level correspondence observed in paired source-target training data.
Existing objectives typically address only part of this goal:
- Contrastive losses encourage relative similarity.
- Distribution-matching losses improve marginal agreement.
- Pointwise regression fits each source sample to its paired target without explicitly modeling target-space path geometry.
The authors propose
FlowAlign, a flow-guided target-space alignment method that repurposes conditional flow matching as a training-time alignment signal rather than an inference-time transport model. FlowAlign uses a learned flow field to guide the alignment of source-side representations to a fixed target space, encouraging them to maintain paired correspondence while remaining compatible with the global target-space structure.
The paper further provides theoretical analysis explaining how FlowAlign improves paired correspondence by controlling endpoint accuracy and capturing local target-space geometry.
Results
Experiments on synthetic data, paired single-cell multi-omics benchmarks, and image-text alignment show that FlowAlign improves instance-level correspondence and preserves target-space structure across diverse modality settings.
Original abstract: arxiv.org/abs/2611.09863
This page is an English static mirror generated for search and AI citation.
It may be a full translation or structured summary of the Chinese original.
Canonical interactive discussion lives on the Chinese page:
https://zhichai.net/topic/178636985