SlerpFlow: Spherical Trajectory Correction for Rectified Flow Inversion

Rectified-flow-based diffusion transformers, specifically FLUX, excel in high-quality image generation. However, fast and accurate inversion for faithful reconstruction and editing is hindered by discretization errors in linear solvers. To address this, SlerpFlow is introduced as a straightforward zero-shot approach that maximizes FLUX potential for high-fidelity inversion. Unlike RF-Solver, which utilizes complex numerical approximations like high-order Taylor expansions to correct trajectory errors, SlerpFlow provides a more effective method for transforming images back to latent noise.