This repo explores various applications of Flow Matching.
Originally the repo was meant to give intuition for how Flow Matching models can be implemented in a simple way. The repo has moved on from this, but an implementation of a simple conditional flow matching objective can still be found at the branch simple_flow_matching
I also wrote an article explaining how Flow Matching works: https://medium.com/@uisdahl/understanding-flow-matching-de2f706cb09d
- Original Flow Matching paper: https://arxiv.org/abs/2210.02747
- Mini-batch Optimal Transport Flow Matching paper: https://arxiv.org/abs/2302.00482
- VQ-VAE paper: https://arxiv.org/abs/1711.00937
The flow matching model can be found in u_net
Generating flows that take samples from a standard gaussian to the flower-102 distribution
The learned flows create decent samples after around 20 epochs, but there is some collapse in the distribution of colors