FlowMatchingScheduler<T>

Models & Types · AiDotNet.Diffusion.Schedulers

Flow matching scheduler implementing rectified flow ODE sampling.

Inheritance: NoiseSchedulerBase<T> → FlowMatchingScheduler<T>

For Beginners

Flow matching works like drawing a straight line from noise to image.

Traditional diffusion (DDPM/DDIM):

  • Adds noise using: x_t = sqrt(alpha) * image + sqrt(1-alpha) * noise
  • The path from noise to image is curved (follows a complex schedule)
  • Needs many steps to follow the curved path accurately

Flow matching (rectified flow):

  • Uses simple linear interpolation: x_t = (1-t) * image + t * noise
  • The path from noise to image is a straight line
  • Can traverse the straight path in fewer steps (often 20-50 vs 50-100+)
  • The model predicts velocity v = noise - image (direction to move)

ODE sampling step:

  • Start at x_1 (pure noise)
  • At each step: x_{t-dt} = x_t - dt * v(x_t, t)
  • After all steps: arrive at x_0 (clean image)

Used by:

  • Stable Diffusion 3 (SD3)
  • FLUX.1
  • Stable Diffusion 3.5

How It Works

Flow matching is a fundamentally different approach from DDPM-style diffusion. Instead of learning to predict noise (epsilon) at each step, the model learns a velocity field v(x_t, t) that defines an ordinary differential equation (ODE) transporting data between noise and signal.

Reference: Lipman et al., “Flow Matching for Generative Modeling”, ICLR 2023; Liu et al., “Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow”, ICLR 2023

Constructors

Constructor Summary
FlowMatchingScheduler(SchedulerConfig<T> config) Initializes a new flow matching scheduler with rectified flow defaults.

Methods

Method Summary
Dictionary<string, object> GetState()
FlowMatchingScheduler<T> CreateDefault() Creates a flow matching scheduler with default SD3/FLUX configuration.
Vector<T> AddNoise(Vector<T> originalSample, Vector<T> noise, int timestep) Adds noise using flow matching linear interpolation.
Vector<T> Step(Vector<T> modelOutput, int timestep, Vector<T> sample, T eta, Vector<T> noise) Performs one flow matching ODE step (Euler method).
void SetTimesteps(int inferenceSteps) Sets up linearly-spaced inference timesteps for flow matching.