Rectified Flow Generative Models
Rectified flow is a generative modeling technique that learns a velocity field transporting samples from a noise distribution to a data distribution along nearly straight-line ODE trajectories, trained by regressing the constant velocity of the linear interpolation between paired noise and data samples. Because the learned transport is close to straight, high-fidelity samples can be produced in very few solver steps, unlike the many-step stochastic reverse process of classical diffusion. It is the continuous-time, ODE-based successor to score-based diffusion models.
Rectified Flow Training in Stable Diffusion Using MLP Backbones and ODE Solvers
Rectified Flow Training utilizes a rectification principle where conditional ODE solvers predict the drift force between base and target distributions to generate continuous trajectories along shorte…