SlerpFlow: Spherical Trajectory Correction for Rectified Flow Inversion

Researchers introduced SlerpFlow, a zero-shot geometry-based method to improve rectified-flow inversion for image generation.

It replaces more complex trajectory corrections with spherical linear interpolation in latent space, while caching corrected velocities to preserve first-order efficiency.

Experiments on FLUX models show better reconstruction fidelity and more aligned semantic editing, with reduced discretization error and no extra training.