Flow Matching

Flow Matching for Generative Modeling

Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, Matt Le · ICLR 2023

Read paper · arXiv:2210.02747

Learn a continuous generative flow by regressing the velocity field of conditional probability paths.

What you will implement

Implement the conditional path, target velocity, and CFM regression loss; time sampling, Euler integration, and training helpers are supplied.

• 10. Conditional Path and Velocity Regression

  Implement the linear conditional path, target velocity, and conditional Flow Matching loss.

  Implement interpolate, cfm_target, and cfm_loss. Time sampling, the Euler solver, and the test training loop are provided.