Pytorch Conditional-Flow-Matching(CFM) Tutorial

A simple tutorial of Conditional Flow Matching Models (CFMs, Y. Lipman et. al., 2022) using MNIST dataset.

Implementations

• Conditional Flow Matching
• MeanFlow

Prerequisites

(1) Download Pytorch and etcs.

• Anaconda
• Python=3.11
• Pytorch=2.5. with CUDA 12.4

(2) Install dependencies via following command

sh install.sh

Expremental Results

• Used a RTX-3090 GPU for all implementations.

• trained on MNIST dataset for 200 epochs

• ground-truth samples
  

• generated samples
  

References

[1] Neural Ordinary Differential Equations, R. T. Q. Chen et. al., 2018

[2] Denoising Diffusion Probabilistic Models, J. Ho et. al., 2020

[3] Flow Matching for Generative Modeling, Y. Lipman et. al., 2022

[4] Mean Flows for One-step Generative Modeling, Z. Geng et. al., 2025