CME 296: Diffusion & Large Vision Models (Stanford, Spring 2026)
by Stanford University (Afshine Amidi and Shervine Amidi)
Stanford's eight-lecture course on how diffusion, score matching and flow matching image models are built, trained and evaluated.
Overview
CME 296: Diffusion & Large Vision Models is a Stanford course taught in Spring 2026 by Afshine Amidi and Shervine Amidi, the twin instructors behind CME 295 (Transformers & Large Language Models) and the widely shared Stanford ML cheatsheets. The lectures ran on Fridays from April 3 to May 29, 2026, were recorded, and every video, slide deck, midterm and final exam (with solutions) is linked from the public syllabus. There is no homework; assessment was two exams. The eight lectures move in order. Lecture 1 covers DDPM diffusion, the ELBO training derivation and faster sampling with DDIM. Lecture 2 covers score matching, denoising score matching, the SDE formulation and probability flows. Lecture 3 covers flow matching, conditional flow matching and rectified flow, and ties all three families together. Lecture 4 covers VAEs, latent diffusion, CLIP and SigLIP representations, and classifier-based and classifier-free guidance. Lecture 5 covers U-Nets, Diffusion Transformers and multimodal DiT. Lecture 6 covers the training lifecycle: pretraining, supervised finetuning, preference tuning with Flow-GRPO and Diffusion-DPO, DreamBooth, LoRA and distillation. Lecture 7 covers evaluation, from human ratings and Elo to FID, CLIPScore, LPIPS and MLLM-as-a-judge. Lecture 8 extends to video and compares the image and text worlds. I Programmer reported in June 2026 that the full materials, previously part of a paid online offering, are now free.
At a Glance
- Topic
- Models
- Level
- Advanced
- Format
- Course
- Cost
- Free
- Duration
- 8 recorded lectures (each a 110-minute Friday class slot), plus a midterm and final with posted solutions; self-paced
- Provider
- Stanford University (Afshine Amidi and Shervine Amidi)
- Hands-on
- No
- Certificate
- None
What You’ll Learn
- ✓Derive the DDPM diffusion training objective from the ELBO and sample faster with DDIM
- ✓Explain score matching, denoising score matching and the SDE view of diffusion with probability flows
- ✓Train with conditional flow matching and rectified flow, and relate them to diffusion and score models
- ✓Use VAEs, latent diffusion, CLIP and SigLIP embeddings, and classifier-free guidance for text-to-image control
- ✓Compare U-Net, Diffusion Transformer and multimodal DiT architectures and their optimizations
- ✓Tune image models with supervised finetuning, Flow-GRPO, Diffusion-DPO, DreamBooth, LoRA and distillation
- ✓Evaluate generators with human ratings, Elo, FID, CLIPScore, PickScore, LPIPS and MLLM-as-a-judge
- ✓Extend image generation ideas to video and map them onto techniques from the language model world
Highlights
- •Every lecture has a public YouTube recording and slide PDF linked from the syllabus page
- •Midterm and final exams are posted with full solutions, so you can test yourself against the Stanford bar
- •Covers diffusion, score matching and flow matching side by side, then shows how the three families connect
- •Includes post-training for image models (Flow-GRPO, Diffusion-DPO, LoRA, distillation) and evaluation, which most diffusion tutorials skip
- •Taught by the authors of CME 295 and the Stanford ML cheatsheets, so it pairs with their LLM course
Who It’s For
Best For
- ✓ML engineers moving from LLM work into image or video generation
- ✓Engineers fine-tuning or serving Stable Diffusion, Flux or DiT-style models who want the theory underneath
- ✓Researchers and graduate students who want a compact, current survey of diffusion and flow matching
- ✓Readers who finished CME 295 and want the vision-side companion course
Prerequisites
- •Linear algebra, probability and multivariable calculus (the course states this background is ideal)
- •Working knowledge of machine learning and neural network training
- •Comfort following mathematical derivations such as the ELBO and stochastic differential equations
FAQ
What is CME 296: Diffusion & Large Vision Models (Stanford, Spring 2026)?
CME 296 is Stanford's Spring 2026 course on diffusion-based generative models for vision, taught by Afshine and Shervine Amidi. It is for engineers and researchers with linear algebra, probability, calculus and machine learning behind them who want to understand how modern image generators such as Diffusion Transformers work, how they are trained and tuned, and how their output is judged.
Is CME 296: Diffusion & Large Vision Models (Stanford, Spring 2026) free?
CME 296: Diffusion & Large Vision Models (Stanford, Spring 2026) is free to access.
What level is CME 296: Diffusion & Large Vision Models (Stanford, Spring 2026) for?
CME 296: Diffusion & Large Vision Models (Stanford, Spring 2026) is aimed at a advanced audience. Recommended background: Linear algebra, probability and multivariable calculus (the course states this background is ideal), Working knowledge of machine learning and neural network training, Comfort following mathematical derivations such as the ELBO and stochastic differential equations.
How long does CME 296: Diffusion & Large Vision Models (Stanford, Spring 2026) take?
Expect roughly 8 recorded lectures (each a 110-minute Friday class slot), plus a midterm and final with posted solutions; self-paced. Most learners work through it at their own pace.
What will I learn from CME 296: Diffusion & Large Vision Models (Stanford, Spring 2026)?
You'll learn: Derive the DDPM diffusion training objective from the ELBO and sample faster with DDIM; Explain score matching, denoising score matching and the SDE view of diffusion with probability flows; Train with conditional flow matching and rectified flow, and relate them to diffusion and score models; Use VAEs, latent diffusion, CLIP and SigLIP embeddings, and classifier-free guidance for text-to-image control; Compare U-Net, Diffusion Transformer and multimodal DiT architectures and their optimizations; Tune image models with supervised finetuning, Flow-GRPO, Diffusion-DPO, DreamBooth, LoRA and distillation; Evaluate generators with human ratings, Elo, FID, CLIPScore, PickScore, LPIPS and MLLM-as-a-judge; Extend image generation ideas to video and map them onto techniques from the language model world.
Topics
Sources
This page was written from 3 sources, 1 on domains other than cme296.stanford.edu.