Flow Maps: Accelerating Diffusion Model Sampling by Learning the Integral Directly

Summary

This article explores flow maps as an alternative to iterative sampling in diffusion models. Instead of taking many small steps to denoise and generate samples (calculating an integral across noise levels), the author proposes training neural networks to directly predict this integral, which could significantly speed up sampling. The article provides a deep dive into the mathematical foundations of diffusion models, the concept of integrating along the denoising path, and how flow maps offer a more efficient approach by learning the entire transformation in one shot.