FlowAlign
FlowAlign is a target-space alignment method based on conditional flow matching that repurposes flow matching from an inference-time transport model into a training-time alignment signal, using the path geometry provided by the flow field to jointly optimize instance-level correspondence and global structure preservation.
Timeline (last 90 days)
Theoretical analysis of FlowAlign shows that the method improves paired correspondence by controlling endpoint accuracy and capturing the local geometry of the target space.
The paper conducts experiments on synthetic data, paired single-cell multi-omics benchmarks, and image-text alignment tasks, and the results show that FlowAlign improves instance-level correspondence under different modality settings while preserving target-space structure.
In single-cell multi-omics data, FlowAlign can better align cell representations across omics modalities, aiding downstream cell type annotation and trajectory inference.
In image-text alignment, FlowAlign improves the performance of cross-modal retrieval.
FlowAlign is expected in the future to extend to more modalities and more complex downstream tasks, such as multimodal generation and cross-modal reasoning.
All Facts (5)
Theoretical analysis of FlowAlign shows that the method improves paired correspondence by controlling endpoint accuracy and capturing the local geometry of the target space.
50%UnverifiedThe paper conducts experiments on synthetic data, paired single-cell multi-omics benchmarks, and image-text alignment tasks, and the results show that FlowAlign improves instance-level correspondence under different modality settings while preserving target-space structure.
50%UnverifiedIn single-cell multi-omics data, FlowAlign can better align cell representations across omics modalities, aiding downstream cell type annotation and trajectory inference.
50%UnverifiedIn image-text alignment, FlowAlign improves the performance of cross-modal retrieval.
50%UnverifiedFlowAlign is expected in the future to extend to more modalities and more complex downstream tasks, such as multimodal generation and cross-modal reasoning.
50%