Research map / Probabilistic and scientific modeling / Variational and generative sampling
Switched Flow Matching: Eliminating Singularities via Switching ODEs
Read the paperCheck your idea against 108,967 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 Switched Flow Matching (SFM), which samples a discrete switching signal to pick among multiple conditional ODEs, eliminating the singularities that a single flow-matching ODE cannot handle, and reports CIFAR-10 FID results against CFM variants.
Problem
Single-ODE flow matching models suffer singularities when source or target distributions are heterogeneous (multi-modal or disconnected), making optimal transport flows discontinuous and hurting training and sampling efficiency (many NFEs).
Approach
Model source and target as mixtures of conditional distributions indexed by a switching signal s; each s selects its own conditional ODE, with marginal vector field obtained by marginalizing over s. This allows trajectories from different ODEs to intersect and avoids singularities. Compatible with independent and minibatch OT couplings within each switch.
Key results
- On CIFAR-10 with adaptive step size, I-SFM (two2ten, mixed) reaches FID 4.27 and I-SFM (two2ten, extremal) reaches 4.22, versus 4.91 for OT-CFM.
- At NFE 10 on CIFAR-10, I-SFM (two2ten, mixed) achieves FID 49.53 versus 76.41 for OT-CFM, a substantial low-NFE gain.
- Standard I-CFM performs poorly (FID 94.55 adaptive) on CIFAR-10 with a two-mode Gaussian mixture source, while SFM variants that separate source modes outperform OT-CFM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| switch from I-CFM to two-to-ten mixed I-SFM | CIFAR-10 generation · CIFAR-10, NFE 10 | FID: $122.44$ -> $49.53$ | Helped |
| switch from OT-CFM to two-to-ten mixed I-SFM | CIFAR-10 generation · CIFAR-10, NFE 10 | FID: $76.41$ -> $49.53$ | Helped |
| use one-to-ten switching rather than two-to-ten mixed switching (ablation) | I-SFM · CIFAR-10, NFE 10 | FID: $49.53$ -> $93.48$ | Hurt |
| use one-to-ten switching rather than two-to-ten mixed switching (ablation) | OT-SFM · CIFAR-10, NFE 10 | FID: $49.69$ -> $93.04$ | Hurt |
| use OT coupling instead of independent coupling with two-to-ten mixed switching (ablation) | SFM · CIFAR-10, NFE 10 | FID: $49.53$ -> $49.69$ | Hurt |
| use two-to-one switching rather than two-to-ten mixed switching (ablation) | I-SFM · CIFAR-10, NFE 10 | FID: $49.53$ -> $78.46$ | Hurt |
| use two-to-one switching rather than two-to-ten mixed switching (ablation) | OT-SFM · CIFAR-10, NFE 10 | FID: $49.69$ -> $84.32$ | Hurt |
| use extremal instead of mixed two-to-ten switching (ablation) | I-SFM · CIFAR-10, NFE 10 | FID: $49.53$ -> $50.12$ | Hurt |
| use extremal instead of mixed two-to-ten switching (ablation) | OT-SFM · CIFAR-10, NFE 10 | FID: $49.69$ -> $58.25$ | Hurt |
Surprising findings
- Single-ODE CFM cannot transport a Dirac mass to a two-point target under optimal coupling, a limitation SFM removes.
- OT-SFM underperformed I-SFM in several CIFAR-10 settings, which the authors attribute to switching already alleviating singularities.
Limitations
- Identifying singularities in high-dimensional image data is left as future work.
- Gaussian noise level sigma is set to 0 in all experiments, leaving its effect unexplored.
- Source distribution for CIFAR-10 is a two-mode Gaussian mixture rather than the standard Gaussian, so comparisons to prior standard-setting results are not direct.
Experimental setup
Datasets: CIFAR-10 2-d Gaussian mixtures (synthetic) checkerboard (synthetic)
Benchmarks: CIFAR-10
Metrics: FID
Baselines: I-CFM OT-CFM
Builds on: flow matching conditional flow matching rectified flow minibatch optimal transport continuous normalizing flow
Closest papers by meaning
- Optimal Flow Matching: Learning Straight Trajectories in Just One Step (NeurIPS 2024 · similarity 0.90)
- Simulation-Free Training of Neural ODEs on Paired Data (NeurIPS 2024 · similarity 0.90)
- Diff2Flow: Training Flow Matching Models via Diffusion Model Alignment (CVPR 2025 · similarity 0.88)
- Improving Flow Matching by Aligning Flow Divergence (ICML 2025 · similarity 0.88)
- Elucidating Flow Matching ODE Dynamics via Data Geometry and Denoisers (ICML 2025 · similarity 0.88)
- Simple ReFlow: Improved Techniques for Fast Flow Models (ICLR 2025 · similarity 0.88)
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.