Stanford CME296 Diffusion & Large Vision Models | Spring 2026 | Lecture 3 - Flow matching
TL;DR
This lecture introduces flow matching as a third paradigm for generative modeling, explaining how it deterministically transports probability distributions from Gaussian noise to data through learned vector fields, while contrasting its velocity-based mechanics with diffusion and score matching approaches.
🎯 Core Concepts & Conventions 3 insights
Reversed temporal indexing
Unlike diffusion models where t=0 represents clean data, flow matching conventions define t=0 as the initial Gaussian noise distribution and t=1 as the target data distribution.
Probability transport objective
The core goal is transporting the entire probability density from an initial distribution P0 to a target distribution P1 while maintaining the conservation of mass property where total probability sums to one.
Flow and probability path
The flow s_t(x_0) represents the mapping from initial conditions to positions at time t, while the probability path P_t(x) describes the evolving intermediate distributions between noise and data.
🚗 Vector Fields vs. Score Functions 3 insights
Vector fields as velocity instructions
The vector field u_t(x) specifies both direction and speed for particles at specific spatial locations and times, functioning like dynamic routing instructions for self-driving cars.
Score functions as compasses
While the score acts as a compass pointing toward high-density regions, the vector field provides explicit velocity vectors that guide the deterministic transport of samples through space.
Trajectories over sampling
Flow matching focuses on continuous trajectories that map individual points from noise to data, contrasting with score matching's emphasis on navigating density gradients.
📐 Mathematical Foundations 3 insights
Governing ordinary differential equation
Particle trajectories follow the ODE dx/dt = u_t(x), where infinitesimal position changes are determined by the vector field multiplied by the time differential dt.
Lipschitz continuity requirement
Unique trajectories for given initial conditions are mathematically guaranteed only when the vector field satisfies Lipschitz continuity, preventing scenarios where multiple paths originate from the same point.
Conservation via continuity equation
The framework ensures probability density evolves according to the continuity equation, balancing the rate of density change against the divergence of the vector field to prevent mass loss during transport.
Bottom Line
Flow matching offers a deterministic alternative to stochastic diffusion by learning a Lipschitz-continuous vector field that transports samples from Gaussian noise to data distributions through continuous probability paths.
More from Stanford Online
View all
Stanford CS547 HCI Seminar | Spring 2026 | The Modern Motivators of Play
The speaker challenges the game industry's outdated assumption that players primarily seek competition, presenting 2024 data showing only 18% of gamers are motivated by competition while 50% seek stress relief and 40% want community. They introduce a framework of nine motivators divided into classic (Fun, Mastery, Competition, Immersion, Meditation, Comfort) and modern (Self-expression, Companionship, Education), arguing that successful games must layer social and creative motivators onto traditional designs to serve contemporary player needs.
Stanford MS&E435 Economics of the AI Supercycle | Spring 2026 | Applications, Applied AI
Base 10 CEO Tuhin explains why AI inference is shifting from frontier models to custom post-trained models as companies scale, driven by 70-90% cost savings, latency requirements, and the strategic need to own proprietary data rather than feed it to potential competitors.
Stanford CS336 Language Modeling from Scratch | Spring 2026 | Guest Lecture: Dan Fu
Dan Fu explains how LLM inference serves as the engine converting electricity into intelligence, detailing the lifecycle of requests through modern serving systems and emphasizing that GPU kernel expertise enables full-stack ML innovation.
Stanford Robotics Seminar ENGR319 | Spring 2026 | Leveraging Geometry in Robot Learning
Rob Platt argues that modern Vision-Language-Action models discard geometric structure, requiring massive datasets to relearn physical constraints. He proposes hybrid approaches that embed geometric symmetries (equivariance) directly into learning architectures, enabling data-efficient robot policies that respect physical laws.