Flow Maps Enable Neural Networks to Predict Diffusion Paths Directly

Flow maps enable neural networks to predict diffusion model paths directly

Flow maps are a technique that allows neural networks to directly predict any point on a diffusion model's path from any other point, rather than estimating tangent directions at each step. This approach can speed up sampling and also enables more efficient reward-based learning and improved sampling steerability.

---

Sampling from a diffusion model is an iterative process. At each step, a denoiser estimates the tangent direction to a path through input space, and the model moves along this path by repeatedly taking small steps, effectively calculating an integral across noise levels. This gradually transforms samples from a simple noise distribution into samples from a target distribution.

Flow maps offer a different approach. Rather than predicting the tangent direction at each point along a path, a flow map is able to predict any point on a path from any other point on that same path. This makes them a promising tool for faster and more efficient sampling from diffusion models.

The technique has recently become a popular subject of study, building on years of research into making diffusion models faster and cheaper to use. Diffusion distillation, a related approach explored previously, has been one of the main tools used to reduce the number of steps required to obtain high-quality samples, and flow maps represent a further development in this space.

How Flow Maps Work

The key to understanding flow maps lies in viewing diffusion models as defining a bijection between noise and data, with unique paths connecting pairs of samples from each distribution in such a way that they never cross each other. Flow maps build directly on this perspective.

While it is relatively straightforward to define what a flow map is, there are many different ways to build and train them. The literature surrounding the topic is noted as being rife with different formalisms and terminology, which can make it difficult to understand how various approaches fit together. A taxonomy proposed by researchers, referred to in the source as Boffi et al., is used as a primary framework for organising these ideas.

Beyond faster sampling, flow maps have additional capabilities. They can enable more efficient reward-based learning and improved sampling steerability, making them useful beyond simply reducing the number of steps needed during inference.

Getting Started with Flow Maps

The subject is explored in depth in a detailed blog post on sander.ai, which aims to clarify the landscape of flow map research and terminology. The post assumes some familiarity with diffusion models and notes that comfort with vector calculus will help when understanding training methods, though other sections are intended to remain accessible without that background.

For those looking to build foundational knowledge, a comprehensive monograph on diffusion models by Chieh-Hsin Lai and colleagues is recommended in the source as both a refresher and a starting point, combining mathematical rigour with intuitive explanations.

Story based on discussion on Hacker News.