CMU 10799 S26: Lecture 4 - Score-based Models - Diffusion & Flow Matching

Lecture recording of Carnegie Mellon University's Spring 2026 Class: 10799 Diffusion & Flow Matching Lecture 4: Score-based Models Taught by Yutong (Kelly) He Class Website: https://kellyyutonghe.github.io/10799S26/ Want to understand how Stable Diffusion, DALL-E, and Sora actually work – and how to build something even better? This course takes you from mathematical foundations to hands-on research frontiers in diffusion models and flow matching, the generative AI frameworks reshaping computer vision and beyond. In this class, you will explore topics from foundational probabilistic modeling through modern advances: denoising diffusion models, score-based SDEs, flow matching, fast sampling algorithms, controllable generation, flow maps & distillation methods, and discrete variants. Choose your path to level up – fidelity (photorealistic quality), controllability (precise user control), or speed (real-time generation) – and build from scratch towards a complete working system through cumulative homework. You’ll strengthen both your theoretical understanding and practical implementation skills by the end of this course. This class has no exams and is ChatGPT friendly! You are free to use resources like pre-trained models, open-sourced GitHub repositories and AGI-powered coding assistants for your assignments!