How AI Creates Images

How can a picture emerge from random noise—and why might it still get a bicycle’s wheels wrong?

Text-to-image AI can make a rainy Parisian street feel almost photographic. But that does not mean the system searched for the right picture or understood the scene as a person would. It learned patterns from images and associated text, then used those patterns to generate a new arrangement.

In How AI Creates Images, Leo Odåsnac follows one scene—a red bicycle outside a blue café—to explain the process from training to pixels. The central idea is diffusion: a model learns to reverse the progressive corruption of images, then uses that ability to move from noise toward an image guided by a prompt. Along the way, discover how words become numerical signals, how compressed visual representations can be decoded into pixels, and why changing the prompt or random seed can change the result.

The book also looks at where generation stumbles: counting, placement, physical plausibility, and words rendered inside an image. It distinguishes diffusion from autoregressive and flow-based approaches, and briefly connects image generation to language, video, and music.

This is an explanation of the technology, not a prompt-writing manual. Read it to replace the magic with a clear mental model—and to understand both what these systems can make and what they can get wrong.