FlowAlign aligns source representations to target space using flow matching
Original title: Flow-Guided Target-Space Alignment via Path Consistency
FlowAlign repurposes conditional flow matching to preserve paired correspondences while matching global target distribution structure.
Multimodal alignment maps source data to a target space for downstream use. Existing methods address only parts of the goal. FlowAlign guides alignment with learned flow fields to maintain instance-level pairs and global structure. Experiments show improved correspondence on single-cell and image-text benchmarks.
- Flow-guided target-space alignment
- Repurposes conditional flow matching for training
- Improves paired correspondence
- Preserves local target-space geometry
- Theoretical analysis on endpoint accuracy
From the original article
1 min read at arxiv.orgMany 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.
By Ruizhi Yuan · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.