Many AI image generators work by a method called diffusion. During training the system learns how images turn into random noise step by step, then learns to reverse that. To create a picture it starts from noise and gradually cleans it up, guided by a text description, until a coherent image appears.
This approach produces varied, detailed images and gives fine control over the result. The same idea now extends to audio and video. Generation can be slow because it takes many steps, so a lot of research aims to reach good results in fewer steps, and the quality still depends heavily on the data the model learned from.