Research map / Image generation / Fast diffusion sampling
Flow-Guided Target-Space Alignment via Path Consistency
Ruizhi Yuan, Zeqiu Yu, Wei Gao, Wei Chen, Xiwei Tang, Lu Tang
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a flow-guided method for aligning a source modality into a fixed target-modality space, preserving target distribution structure and paired instance-level correspondence.
Problem
Multimodal alignment where a source modality must be mapped into a fixed target-defined space that is used directly for downstream analysis, requiring both global target-distribution structure and instance-level source-target correspondence to be preserved.
Approach
Uses flow-based generative modeling together with a path-consistency objective to align source representations to the target space while preserving global target structure and paired instance correspondence; based only on the abstract, details are unknown.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add local-start sampling (sigma > 0) (ablation) | FlowAlign · PBMC | FOSCTTM: 0.063 -> 0.039 | Improved |
| add local-start sampling (sigma > 0) (ablation) | FlowAlign · PBMC | Cell match: 0.082 -> 0.141 | Improved |
| add source conditioning to the velocity field (ablation) | FlowAlign · PBMC | FOSCTTM: 0.055 -> 0.039 | Improved |
| use field as training-time score rather than inference-time CFM transport (ablation) | FlowAlign · PBMC | LTA: 0.892 -> 0.920 | Improved |
| use field as training-time score rather than inference-time CFM transport (ablation) | FlowAlign · PBMC | FOSCTTM: 0.034 (CFM transport) vs 0.039 | Hurt |
| evaluate frozen field along candidate-dependent paths (ablation) | FlowAlign · PBMC | FOSCTTM: 0.047 -> 0.039 | Improved |
| use very large local-start scale sigma (ablation) | FlowAlign · PBMC | alignment and structure metrics | Hurt |
| replace MSE regression with path-consistency loss | FlowAlign · BMMC | FOSCTTM: 0.042 -> 0.019 | Improved |
| add MMD, InfoNCE or neighborhood term to MSE | MSE mapper · PBMC | FOSCTTM: 0.048 -> 0.047 / 0.053 / 0.047 | No clear effect |
| replace MSE with FlowAlign for image-to-text mapping | FlowAlign · CUBICC | LTA (%): 57.49 -> 70.42 | Improved |
| use FlowAlign vs MSE regression | FlowAlign · Synthetic 2D | MAE_a: 5.61e-3 -> 5.10e-3 | Improved |
| use FlowAlign vs MMD alignment | FlowAlign · Synthetic 2D | MMD: MMD align. 6.99e-3 best | Hurt |
Closest papers by meaning
- When Does Flow Matching Help Deterministic Multimodal Prediction: A Spatial Transcriptomics Study (NeurIPS 2026 · similarity 0.87)
- Value Gradient Guidance for Flow Matching Alignment (NeurIPS 2025 · similarity 0.85)
- Propensity Score Alignment of Unpaired Multimodal Data (NeurIPS 2024 · similarity 0.85)
- Structured Coupling for Flow Matching (NeurIPS 2026 · similarity 0.85)
- Blockwise Flow Matching: Improving Flow Matching Models For Efficient High-Quality Generation (NeurIPS 2025 · similarity 0.84)
- Direct Product Flow Matching: Decoupling Radial and Angular Dynamics for Few-Shot Adaptation (NeurIPS 2026 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
