Research map / Probabilistic and scientific modeling / Variational and generative sampling
Co-Evolving Interpolants and Flows via Path-Flow Alignment
Zeyu Michael Li, William X Chen, Xiang Cheng
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 jointly training an endpoint-preserving path network and a flow network with a shared path-flow alignment loss for flow matching.
Problem
Flow matching typically fixes the interpolation path and learns only the velocity field, which may be suboptimal.
Approach
Co-evolve the interpolation path and the flow: the flow matches the path velocity while the path aligns its velocity to the current flow, using one alignment loss with an endpoint-preserving path network.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add path-flow alignment with stochastic regularization to continued fine-tuning | SiT-S · ImageNet-256 (FM) | 50k-FID: 9.022 -> 8.359 | Improved |
| Add path-flow alignment with stochastic regularization to continued fine-tuning | SiT-B · ImageNet-256 (FM) | 50k-FID: 5.057 -> 4.415 | Improved |
| Add path-flow alignment with stochastic regularization to continued fine-tuning | SiT-L · ImageNet-256 (FM) | 50k-FID: 3.390 -> 2.742 | Improved |
| Add path-flow alignment with stochastic regularization to continued fine-tuning | SiT-XL · ImageNet-256 (FM) | 50k-FID: 2.091 -> 1.825 | Improved |
| Add path-flow alignment to model-guidance fine-tuning | SiT-B-MG · ImageNet-256 (MG) | 50k-FID: 7.940 -> 5.585 | Improved |
| Add path-flow alignment to model-guidance fine-tuning | SiT-XL-MG · ImageNet-256 (MG) | 50k-FID: 1.494 -> 1.322 | Improved |
| Remove stochastic regularization (rho_t = 1) from path-flow alignment (ablation) | SiT-S · ImageNet-256 (FM) | 50k-FID | Hurt |
| Add stochastic source perturbation (rho_t < 1) to path-flow alignment (ablation) | SiT-S · ImageNet-256 (FM) | late-time flow divergence / FID | Improved |
| Increase kinetic energy weight beta (with rho_t = 1) (ablation) | SiT-S · ImageNet-256 (FM) | FID late degradation | Mixed |
| Add acceleration regularization to path objective (ablation) | SiT-S · ImageNet-256 (FM) | 50k-FID | No clear effect |
Experimental setup
Builds on: flow matching
Closest papers by meaning
- Structured Coupling for Flow Matching (NeurIPS 2026 · similarity 0.90)
- Value Gradient Guidance for Flow Matching Alignment (NeurIPS 2025 · similarity 0.89)
- Diff2Flow: Training Flow Matching Models via Diffusion Model Alignment (CVPR 2025 · similarity 0.89)
- Optimal Flow Matching: Learning Straight Trajectories in Just One Step (NeurIPS 2024 · similarity 0.89)
- A Theory on Flow Matching with Neural Networks (NeurIPS 2026 · similarity 0.89)
- Metric Flow Matching for Smooth Interpolations on the Data Manifold (NeurIPS 2024 · similarity 0.89)
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.
