Generative AIintermediate13 min · 9 steps
GANs, VAEs and Diffusion
The three architectures behind image generation: two networks in an adversarial game, a probabilistic latent space, and a model that learns to undo noise one step at a time.
dataerrorsignalresponserequestheavy work
ProblemStep 1 of 9
There's no “right answer” for a new image
Training a classifier is easy to grade: the photo is a cat or it isn't. But to generate an image there's no single correct output to compare against — any of millions of cats would do.
So the question every image model must answer is: how do you score “looks real”? Three architectures give three different answers.
InsightGANs train a judge. VAEs learn a compressed map of the data. Diffusion models learn to remove noise.
← Previous lessonPrompting, Fine-Tuning and RLHFNext lesson →Image Generation: Diffusion, Text-to-Image and Editing
Lesson 4 of 8 in Generative AI