The Flow of Creation: A Tour of Flow Matching for Visuals

• Manolo Canales Cuba UFABC
• Vinícius do Carmo Melício UFABC
• João Paulo Gois UFABC

Resumo

Flow Matching has recently emerged as an efficient alternative to the generative method paradigms. Here we aim to provide researchers and practitioners with both the theoretical and the practical aspects of this technique. We delve into the continuous formulation, where a neural network learns a vector field to transform noise into data via an ordinary differential equation, and also explore its discrete counterpart. The paper covers the entire workflow, from the core mathematical concepts and training objectives to sampling procedures, including classifierfree guidance and conditioning generation. By showcasing a diverse range of applications—from image synthesis and human motion generation to computational biology and robotics—this work equips readers with the essential knowledge to apply and innovate with the versatile and computationally efficient flow matching framework.

Palavras-chave: Training, Visualization, Image synthesis, Noise, Neural networks, Transforms, Ordinary differential equations, Vectors, Pattern matching, Robots, Flow Matching, Discrete Generative Models, Deep Learning