Course

How Modern Image and Video Models Work

Teacher-authored course

Learn with teacher-produced videos and their original narration. Your progress belongs to your active learner profile.

Your path through this course

Module 00 — Course introduction

Welcome: the puzzle of making an image1:36

Module 01 — The generative modeling problem

Where does a new image come from?1:07 What does a generative model learn?1:24 Why generation is difficult1:22 Anatomy of a modern image generator1:29 Five questions for any generator1:33

Module 02 — Diffusion from first principles

A training problem we can manufacture0:57 What can be recovered from a noisy image?1:06 Noise schedules and signal-to-noise ratio1:05 What should the network predict?1:05 One diffusion training step1:04 The score hidden in denoising1:04 Sampling is a sequence of model calls1:00 Stochastic paths and deterministic paths1:00

Module 03 — Latent generative models

Why generate in a latent space?1:03 Compress first, generate later1:08 How the autoencoder earns its latent space1:09 What lives in a latent?1:11 Training and sampling in latent space0:57 Latent scaling is part of the model0:57

Module 04 — Transformers for image generation

From a latent grid to image tokens1:03 Self-attention, one token at a time0:59 What a transformer block adds1:02 Time is an input, not a clock1:00 Text tokens and cross-attention0:58 Joint attention and multimodal streams1:01 Position, aspect ratio, and token count0:59 Trace one complete model call1:04

Module 05 — Flow matching and sampling

From denoising to transport0:59 Draw a path between noise and data1:20 The velocity the network can actually learn0:59 The flow-matching objective1:29 Why generation becomes an ODE1:03 Euler's method in one line1:15 Where diffusion and flow matching meet1:04 Which times does the model practice?0:59 A complete flow model in one pass1:05

Module 06 — Conditioning and control

How a prompt changes the prediction1:05 Classifier-free guidance, derived1:09 What guidance gains and costs1:02 Conditioning on an image1:18 Identity-preserved generation1:12 Inpainting makes constraints explicit1:02

Module 07 — Training a real image model

The dataset defines the task1:04 Quality versus quantity1:15 Captioning and recaptioning1:09 Resolution and aspect-ratio buckets1:10 Build one exact training batch1:12 When to adapt and when to start over1:07 LoRA and adaptation1:26 What distributed training changes1:04 Read a training run before trusting it0:53

Module 08 — Evaluating and debugging image models

Why a lower loss may not look better1:05 Quality and prompt adherence1:09 What FID and embedding scores measure1:01 A human comparison that can be trusted0:58 Vision-language models as judges1:05 Build an evaluation suite from failure modes1:01 Debug by locating the first failure1:00

Module 09 — Stable Diffusion 3 case study

Stable Diffusion 3 as a complete system1:06 What the three text encoders contribute1:09 Inside one multimodal diffusion-transformer block1:04 The actual flow-training target and timestep choice1:11 How the image model was trained at scale1:06 Sampling, evidence, and what to inspect1:15

Module 10 — Video foundations

What time adds to generation0:57 A video latent has three axes of structure1:01 The video autoencoder sets the motion ceiling1:00 Turn a space-time volume into tokens1:01 Attention across space and time1:01 Video data1:07 What image pretraining gives video0:58 Train a flow model over an entire clip0:59 Text-to-video1:10 Image-to-video1:21 Motion and camera control1:09 Evaluating video1:16 Design a video model on a whiteboard0:59

Module 11 — Wan video case study

Wan: put the video pipeline on the board0:57 Why Wan has a causal video autoencoder1:07 Inside Wan’s video transformer1:03 Wan’s flow objective and staged training1:05 Data and captions are part of the architecture story0:59 Evaluate the model, then transfer the method1:07

Module 12 — The next questions

Where the red boat came from1:47 The next questions1:55