Rectified flow
Algorithms AIMech:RectifiedFlow DRAFT
Aliases: RF (rectified flow)
A transport-learning method that regresses a velocity field on endpoint differences along straight interpolations of coupled samples and uses the resulting ordinary differential equation to transport the source distribution, optionally repeating rectification using the learned endpoint coupling.
Identity Notes & Concept Scope
The ignored-inclusive identity search covered both checkouts, all 442 AIO labels and scoped synonyms, and 90 preceding native identities. This record denotes the straight-interpolation regression and rectification procedure of Liu, Gong and Liu, including optional reflow. It overlaps with particular conditional flow-matching objectives, but does not denote every probability-path choice in flow matching, a particular trained image generator, or mandatory distillation.
Claims & Grounded Evidence
2 Claims
Mechanism
#generative-mechanism
Given coupled endpoints X0 and X1, training samples t and regresses v((1−t)X0+tX1,t) toward X1−X0 using squared error. The ideal minimizer is the conditional mean endpoint difference at the interpolated state. Integrating dZt/dt=v(Zt,t) from the source sample induces a new endpoint coupling. Reflow optionally trains another rectified flow on that coupling; endpoint-map distillation is a separate optional step.
Limitation
#optimal-transport-boundary
A later analysis of gradient-constrained rectification exhibits nonoptimal zero-loss fixed points with disconnected intermediate supports; its corrected equivalence theorem additionally requires full intermediate support, rectifiability and smooth potential fields. Straightness or imposing a potential alone does not establish globally optimal transport.
⚠️ Curated Limitations & Negative Results
- Marginal preservation and transport-cost statements concern ideal regression with a well-posed ODE, not arbitrary finite neural estimates or numerical solvers.
- Straight training interpolations do not make the learned marginal trajectories exactly straight; one-step accuracy is not universal. Reflow and distillation are optional.
- The original paper explicitly does not guarantee a globally cost-optimal coupling in general. Later analysis also gives nonoptimal zero-loss fixed points for a gradient-constrained rectification variant when required support or regularity assumptions are absent.
- The later optimal-transport counterexamples limit stronger equivalence claims; they do not refute the original least-squares construction. No executable solver, numerical benchmark or full proof verification is supplied.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:RegressionAnalysis
Regression Analysis
|
related_match | The defining velocity-learning objective estimates conditional endpoint differences by least-squares regression. RegressionAnalysis is the related estimation task, while rectified flow is the transport-learning algorithm and may include iterative reflow; no task/algorithm subclass assertion is made. |
Evidence Sources & Primary Literature
2 Sources| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
original-paper
|
Xingchao Liu, Chengyue Gong, Qiang Liu
|
primary paper
full_text
|
ID: arXiv:2209.03003
sha256:d81790cb6988…
|
optimal-transport-analysis
|
Johannes Hertrich, Antonin Chambolle, Julie Delon
|
primary paper
full_text
|
ID: arXiv:2505.19712
sha256:8aca06c93927…
|
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/RectifiedFlow.yaml