Reproduce “Delta Rectified Flow Sampling for Text-to-Image Editing” 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
Gaspard Beaudouin, Minghan Li, Jaeyeon Kim, Sung-Hoon Yoon, Mengyu Wang
Researchers at Harvard University developed Delta Rectified Flow Sampling (DRFS), an inversion-free method for text-to-image editing that addresses over-smoothing artifacts by formulating an energy function based on velocity residual differences and introducing a time-dependent shift term. DRFS achieves superior semantic alignment and structure preservation on benchmarks while maintaining higher visual fidelity compared to previous distillation sampling techniques.
Key insights
- Minimizing the difference between target and source velocity residuals effectively mitigates over-smoothing by concentrating gradients only in areas requiring change.
- A time-dependent shift term (c_t) that vanishes at the beginning and end of the trajectory, such as c_t textasciitilde (1-t)t, is crucial for balancing strong semantic editing with high fidelity and preventing error amplification.
- DRFS provides a unified theoretical framework, demonstrating equivalence to Delta Denoising Score (DDS) when c_t=0 and recovering FlowEdit's dynamics under specific rectified flow parameterization and shift scheduling.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.