title: Flow Matching: Conditional Vector-Field Training description: Flow matching trains a continuous vector field along a chosen probability path; scheduler, path, and inference settings are checkpoint-specific rather than universal diffusion defaults. category: architectures tags: [flow-matching, conditional-flow-matching, scheduler, sampling, continuous-normalizing-flow, ode] aliases: ["Flow Matching", "Rectified Flow", "Flow Matching Scheduler"]
Flow Matching: Conditional Vector-Field Training¶
Flow Matching for Generative Modeling defines a simulation-free way to train continuous normalizing flows (CNFs). Instead of requiring simulation of the learned flow during training, the model regresses a vector field on a chosen family of conditional probability paths.
The path matters. The original formulation can cover diffusion paths and optimal-transport-style paths; it is not a single universal straight-line interpolation or one fixed equation for every checkpoint.
What inference integrates¶
At inference, a compatible pipeline starts from its configured source distribution and numerically integrates the learned vector field toward a sample. The solver, time schedule, conditioning, guidance, and number of function evaluations are part of that pipeline/model contract.
Do not infer that every flow-matching checkpoint:
- accepts the same scheduler class or schedule shift;
- improves at a particular number of steps;
- supports a particular img2img strength mapping; or
- can safely refine its own output in repeated passes.
Those claims need evidence from the checkpoint publisher and a result on the intended task.
Scheduler compatibility¶
Diffusers provides FlowMatchEulerDiscreteScheduler as one implementation. Its exposed configuration includes training timesteps, shift behavior, dynamic shifting, sequence-length settings, and alternative sigma schedules. These options describe an implementation interface; they are not universal defaults to copy into another flow model.
Load the scheduler and configuration published with the checkpoint whenever possible. If an experiment changes the scheduler, shift, solver, or step schedule, treat it as a new sampling contract and retain a reproducible result before using it operationally.
Evaluation protocol¶
For each supported model/pipeline pair:
- record checkpoint, pipeline and scheduler revisions, input fixture, seed, dimensions, guidance, steps, and device/runtime;
- make a small supported grid of step counts or schedule settings;
- compare output determinism, task-relevant fidelity, artifacts, and latency;
- review img2img or editing preservation separately from text-to-image quality; and
- publish the selected configuration with its validation fixture and known limits.
This is especially important for restoration and editing: an image can look more polished while drifting from the source, changing text, or altering a material property that must be preserved.