GMFlow for Inverse Problems
Variational and Gaussian-mixture flow matching for image inverse problems, validated on 2D toy examples.
Undergraduate research at ZJUI, advised by Prof. Xiangming Meng, Jun 2025 – Dec 2025.
In this project I studied Variational Flow Matching and Gaussian Mixture Flow Matching (GMFlow) for image inverse problems. GMFlow does not predict a single velocity. Instead, it models the velocity distribution as a Gaussian mixture, which can represent multi-modal posteriors. I implemented 2D toy examples to test our new algorithm before scaling it to images.