How does text-to-image AI work?

Many image models are diffusion models. They start from random noise and remove it step by step until an image appears, guided at each step by your text. The method was described in the 2020 paper Denoising Diffusion Probabilistic Models. Providers do not always publish how their current models are built.

Because generation starts from randomness, the same prompt usually gives a different image each time. Some APIs let you fix the random starting value, called a seed, to repeat a result. See Does AI generate the same image twice from the same prompt?.

Newer models also take images as input next to text. That is what makes editing an existing picture with words possible: Google describes its Gemini image models as generating and processing images conversationally with text, images or both.

ImageQuarry turns this into a conversation: you describe an image, it generates it, and you keep asking for changes to the version you choose.