How do AI image generators actually know what a cat looks like?

[Dark Elizabeth]

Genuinely curious how an AI image generator can create a picture of something it's never technically seen, how does it actually know what a cat looks like?

[Coastal Estuary]

It was trained on an enormous number of existing images that were each paired with a text description, millions of pictures labeled cat, dog, sunset, and so on. During training, the model gradually learns statistical patterns, what visual features, shapes, colors and textures tend to show up whenever the word cat appears alongside an image

It doesn't store or recall an actual photo of a cat the way a search engine would, instead it learns an abstract representation of catness, the general visual pattern that makes something recognizable as a cat, built from patterns across millions of examples rather than copied from any single one

When you ask it to generate a cat, it essentially starts with random visual noise and gradually refines that noise step by step, guided by its learned understanding of what a cat pattern looks like, until a coherent image emerges. This is why generated images can look almost right but slightly off, extra fingers, strange ears, since the model is generating from a learned statistical pattern rather than copying a real reference

[Skibidi]

The random noise gradually refined into an image explanation is such a good way to describe diffusion models, way clearer than most technical explanations I've read

[Candle]

This is exactly why weird anatomical mistakes happen so often, it's working from a statistical blend of patterns rather than an actual understood structure

[VacantTundra]

Good clarification that it's not storing actual photos, that misconception comes up constantly in copyright discussions about these tools