01Overview
In 2014 Goodfellow proposed generative adversarial networks: a generator fakes and a discriminator judges, and both improve through the contest. GANs produce convincing images but are hard to train and prone to mode collapse. That same year variational autoencoders (VAEs) offered another route through probabilistic latent variables.
In 2020 denoising diffusion models (DDPM) split generation into hundreds of steps of adding and removing noise: stable to train, high sample quality. In 2021 CLIP mapped text and images into one space, and DALL·E and Stable Diffusion made “drawing with words” a mass-market tool. Video generation arrived by 2024. The next steps are better control, lower cost and verifiable rights.
02History
- 2014
GANs
Adversarial training generates images.
- 2014
VAEs
Variational inference with latent variables.
- 2020
DDPM
Diffusion models.
- 2021
CLIP and DALL·E
Aligning text and images.
- 2022
Stable Diffusion
Open-source image generation.
03Core concepts
Adversarial training
Generator and discriminator compete and approach the true distribution.
Diffusion process
Add noise, then learn to remove it step by step to sample from noise.
Multimodal alignment
Text and images correspond inside one semantic space.
04Hands-on
Do it once — it beats reading it ten times
Train a tiny VAE on MNIST and look at reconstructions versus samples — explain the difference
05Self-check
Say it in your own words first, then open the answer. Whatever you cannot explain is what to review next.
Can you explain "Adversarial training"?
Generator and discriminator compete and approach the true distribution.
Can you explain "Diffusion process"?
Add noise, then learn to remove it step by step to sample from noise.
Can you explain "Multimodal alignment"?
Text and images correspond inside one semantic space.
What problem did it solve, and why was there nothing better at the time?
Go back to the history section: how was this done before, and what did it cost?
If you had to build it today, how would you break it down?
Use the hands-on task above: write a plan of at most three steps before you start.
06Key people
Ian Goodfellow
1987 –
Inventor of GANs.
Robin Rombach
1991 –
Core author of Stable Diffusion.
07Further reading
- Generative Adversarial Networks
- Denoising Diffusion Probabilistic Models