跳到主要内容Skip to main content

Flow MatchingFlow Matching

Flow Matching

  • 图像语音视频Image, speech & video
  • 第六站 · 选得对Station 6 · Choosing Right
  • 引用核验 2026-11-18refs checked 2026-11-18
解决什么问题What problem it solves

学一个从噪声到数据的连续速度场,用常微分方程把噪声推成样本。Learn a continuous velocity field from noise to data and push noise into samples with an ODE.

这是什么What it is

训练时让模型回归「从噪声点指向数据点」的速度,而不是直接预测噪声或像素。

推理时从噪声出发,沿速度场积分若干步,得到样本。

它是扩散模型之外的连续生成框架;Stable Diffusion 3、Flux 等图像模型采用了 rectified flow 变体。

Training regresses the velocity from a noise point toward a data point, rather than predicting noise or pixels directly.

Inference starts from noise and integrates the velocity field for a number of steps.

A continuous generative framework beside diffusion; Stable Diffusion 3 and Flux use rectified-flow variants.

为什么需要知道Why it matters

扩散模型要很多步采样,Flow Matching 把「走哪条路」学得更直,步数可以更少;理解它,才能看懂 2025–2026 图像 / 视频生成模型为什么普遍从「扩散」转向「流」。Diffusion needs many sampling steps; Flow Matching learns a straighter path, so fewer steps can suffice. It explains why image and video generators moved from “diffusion” to “flow” in 2025–2026.

出处Sources
  • Flow Matching(Lipman 2022)Flow Matching for Generative Modeling (2022) ↗
  • Rectified Flow(2022)Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow (2022) ↗