Rectified Flow

Rectified flow learns a time-dependent velocity field that transports simple noise distributions into data, often along straighter paths than diffusion models. Researchers improve these transports through coupling and source-distribution design, consistency training and distillation, and guidance or geometric constraints for tasks such as image editing, restoration, forecasting, and robot control. The approach is being applied across images, audio, medical data, and embodied policies, with work also studying sample complexity, optimal transport, and privacy. A central tension is how to reduce inference to very few steps while preserving fidelity, diversity, and reliable behavior under changing conditions.