Flow Matching for Multimodal Distributions

Twin Cities → Equal Contribution https://mm-flow.github.io Figure 1. (a) The intuition of source and coupling co-design is to reduce crossing (b) when the target distribution exhibits multimodal structures. (c) Our Multimodal Flow Matching (MM-FM) brings 30→ faster convergence for DiT DH -XL trained on DINOv2-B latent space for unconditional generation on ImageNet256 compared with the classic flow matching algorithm.