Reproduce “Variational Rectified Flow Matching” 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
Pengsheng Guo, Alexander G. Schwing
Variational Rectified Flow Matching (VRFM) refines continuous-time generative models by addressing the multi-modality of velocity fields in Rectified Flow Matching. Integrating variational inference, VRFM explicitly models ambiguous flow directions using a latent variable, leading to improved generation quality with superior FID scores on datasets like ImageNet and enhanced controllability of outputs.
Key insights
- Explicitly modeling the multi-modal nature of 'ground-truth' velocity fields in Rectified Flow Matching, instead of averaging them, leads to straighter and more accurate generative flows.
- Integrating a continuous latent variable via a variational inference framework enables the model to represent ambiguous flow directions and disentangle generative factors.
- The proposed variational objective maintains the marginal data distribution, ensuring the theoretical soundness of the generative process.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.