Reproduce “Optimal Transport for Rectified Flow Image Editing: Unifying Inversion-Based and Direct Methods” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Marian Lupascu, Mihai-Sorin Stupariu
By integrating optimal transport theory, this research unifies and enhances both inversion-based and direct methods for rectified flow image editing, achieving improved reconstruction fidelity, structural preservation, and semantic consistency with minimal computational overhead. The framework provides robust, training-free guidance for optimizing editing trajectories in models like FLUX and Stable Diffusion 3.
Key insights
- Optimal Transport theory offers a robust mathematical framework to define and steer optimal trajectories within the latent space of Rectified Flow models, effectively mitigating trajectory deviation during editing.
- The same optimal transport principles can be leveraged to simultaneously improve both the reconstruction fidelity of inversion-based RF methods and the semantic consistency and path optimization of direct RF editing techniques.
- Substantial performance enhancements for RF image editing can be achieved through a training-free integration of OT guidance, maintaining computational efficiency and offering precise control over the fidelity-editability trade-off.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.