Reproduce “Text-to-Image Rectified Flow as Plug-and-Play Priors” 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
Xiaofeng Yang, Cheng Chen, Xulei Yang, Fayao Liu, Guosheng Lin
Research from Nanyang Technological University and A*STAR demonstrates that pretrained text-to-image rectified flow models can function as effective plug-and-play priors for downstream tasks. It introduces Rectified Flow Distillation Sampling (RFDS) and its variants, showing improved text-to-3D generation quality, robust 2D image editing, and faster optimization compared to diffusion-based methods.
Key insights
- Rectified flow models can serve as effective plug-and-play priors for various downstream tasks, often outperforming diffusion-based priors in generation quality and efficiency.
- The time-symmetry of rectified flow is crucial for enabling robust image inversion and subsequent text-guided editing by allowing the optimization of latent noise representations.
- Iteratively refining the latent noise (as in RFDS-Rev) before applying the distillation loss significantly enhances the detail and quality of generated content, addressing common limitations of baseline distillation methods.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.