Learning the integral of a diffusion model
Summary
The article provides a deep dive into flow maps for diffusion models, contrasting stochastic versus deterministic sampling, and detailing how flow maps offer a global path representation between noise and data. It covers the three consistency notions (compositionality, Lagrangian, Eulerian), training regimes (from scratch vs distillation), and practical considerations for implementing flow maps, including mean flow variants and recent extensions.