Reproduce “FlowSlider: Training-Free Continuous Image Editing via Fidelity-Steering Decomposition” on OpenResearch

About this paper

Taichi Endo, Guoqing Hao, Kazuhiko Sumi

FlowSlider introduces a training-free method for continuous image editing in Rectified Flow models by decomposing the velocity update into fidelity and steering components. This approach enables stable slider-style control, preserving source image fidelity while achieving superior performance compared to learning-based and heuristic training-free methods across various editing tasks.

Key insights

• Decomposing the velocity update into fidelity and steering components allows for independent control over semantic change and source preservation, preventing artifacts caused by indiscriminately scaling the entire update.
• The approximate orthogonality between the fidelity and steering vectors ensures that scaling the steering component primarily impacts the edit's magnitude while minimally disturbing the source-preserving dynamics.
• This training-free approach provides a robust and generalizable mechanism for continuous editing in Rectified Flow models, overcoming the limitations of both computationally expensive learning-based methods and unstable heuristic controls.
• The method can inherently support bidirectional editing (e.g., aging and de-aging) by applying negative strength values, offering expanded creative control from a single prompt pair.

Summary by alphaXiv.