Flow Models Exploration
Conditional image generation with flow matching, implemented from scratch on MNIST and CIFAR-10.
Undergraduate research at ZJUI, advised by Prof. Xiangming Meng, Jun 2025 – Dec 2025.
Flow matching trains a neural network to predict a time-dependent velocity field \(u_t(x)\). Integrating this field moves samples from a simple source distribution, usually a standard Gaussian, to the data distribution. In this project I studied the framework in depth and implemented conditional image generation on MNIST and CIFAR-10 from scratch.
The density path \(p_t\) generated by the velocity field satisfies the continuity equation:
\[\frac{\partial p_t(x)}{\partial t} + \nabla \cdot \big(p_t(x)\, u_t(x)\big) = 0 .\]