← All resources

Learning the integral of a diffusion model

Auto-discovered

Blog Advanced

Sander Dieleman

Visit resource ↗

Description

Learning the integral of a diffusion model is a Sander Dieleman blog post on sander.ai explaining flow maps as an alternative parameterization of diffusion and flow-based generative models. Dieleman synthesizes recent research for readers already training diffusion models.

Flow maps predict any point along a noise-to-data trajectory from any other point on that trajectory, not only the local tangent vector field a standard diffusion model learns. The post surveys consistency objectives, training from scratch versus distillation, and applications such as fewer-step sampling and reward-guided finetuning.

Advanced machine learning readers with diffusion background who want a unified map of flow-map papers are the audience. The post is math-heavy but avoids being a beginner diffusion tutorial.

The article is free as a single HTML essay with citations to arXiv work. Dieleman's essays are widely shared in generative model research communities adjacent to neural rendering.