Flow matching
Algorithms AIMech:FlowMatching DRAFT
Aliases: FM (flow matching)
A generative-model training method that fits a time-dependent vector field by regression to velocities associated with a chosen probability path between a base distribution and a data distribution, with conditional flow matching providing a tractable training objective.
Identity Notes & Concept Scope
The 442 pinned AIO identities and 90 preceding native identities, including scoped synonyms and aliases, were compared alongside ignored-inclusive searches of both checkouts. This record denotes the general probability-path training method of Lipman and colleagues. Conditional flow matching is its tractable objective construction, not an asserted synonym for every use of the general method. Straight conditional paths overlap with the training construction used in rectified flow; that overlap does not make the general path family and the rectification/reflow procedure identical.
Claims & Grounded Evidence
1 Claim
Mechanism
#generative-mechanism
The method regresses a neural vector field on a target velocity under a prescribed time-indexed density. Conditional flow matching samples a data-conditioned path and its known conditional velocity; under the paper's assumptions its loss has the same parameter gradient as the marginal flow-matching loss. For the illustrated Gaussian conditional paths, the mean and standard deviation define an affine transformation of base noise. Training avoids integrating the learned ODE, while generation integrates that vector field from the base distribution.
⚠️ Curated Limitations & Negative Results
- The equality of conditional and marginal objective gradients is conditional on the paper's density and integrability assumptions, including positive path density in Theorem 2; finite neural training does not guarantee the ideal field.
- The Gaussian construction with positive terminal noise leaves a smoothed approximation to the data distribution. Training without ODE simulation does not remove numerical ODE solving from sampling.
- The optimal-transport conditional Gaussian path in Section 4 does not imply that the resulting marginal flow is a globally optimal transport map, as the paper explicitly notes on PDF p. 6.
- No benchmark superiority, universal sampling-step count or independently checked theorem proof is asserted.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:RegressionAnalysis
Regression Analysis
|
related_match | Flow matching uses regression of time- and state-dependent velocity targets. RegressionAnalysis describes the corresponding statistical estimation task, whereas this local concept denotes a generative training algorithm; the mapping does not assert an algorithm is a task. |
Evidence Sources & Primary Literature
1 Source| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
original-paper
|
Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, Matt Le
|
primary paper
full_text
|
ID: arXiv:2210.02747
sha256:5eeb39ba5163…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-11T04:42:32Z
Updated: 2026-10-11T04:42:32Z
Source File:
data/curated/FlowMatching.yaml