How does the text prompt control the image?

The prompt is turned into an embedding by a text encoder, and that embedding conditions every denoising step. Wording matters because it moves that vector, which is why a synonym can change the result more than you expect.

Why do generated hands and text still look wrong?

Both need exact, countable structure, and the model optimises for what looks statistically plausible rather than what is structurally correct. Results have improved a lot as models got better at global consistency, and these remain the places errors show first.

Is diffusion related to how ChatGPT works?

They share the transformer machinery and the idea of learning a distribution, and they differ in the loop. A language model predicts the next token in a sequence. A diffusion model refines a whole image over repeated passes. Some newer systems blur that line by using diffusion for text or transformers inside the denoiser.

Plain-language explainer

Diffusion, or how an image appears out of noise

How do AI image generators actually work?

A diffusion model is trained by taking real images, adding noise until they are static, and learning to undo one step of that noise. To generate, it starts from pure random noise and runs the learned undo step many times, each time nudging the picture toward something more image-like. Your text prompt steers every step, so the noise resolves into a picture that matches the description. Nothing is copied or retrieved. The image is denoised into existence.

Last reviewed July 25, 2026

Reading is the slow way. Start with a free lesson you can operate right now.

Start free: From the internet to your answer →

Free, no code, no signup.

Then go deeper: Diffusion: images from noise Locked

What people get wrong

It stitches together pieces of images it has seen. It stores learned patterns in weights, not a library of source images to paste from.
It draws like a person, starting with outlines. It refines the entire canvas at once, dozens of times, from blur to detail.
More steps always means a better image. Past a point extra steps change very little and just cost more time.

Where you see it in real products

Image generators such as Midjourney, DALL-E and Stable Diffusion.
Generative fill and object removal in photo editors.
Product and marketing mockups produced from a written brief.

Common questions

How does the text prompt control the image?: The prompt is turned into an embedding by a text encoder, and that embedding conditions every denoising step. Wording matters because it moves that vector, which is why a synonym can change the result more than you expect.
Why do generated hands and text still look wrong?: Both need exact, countable structure, and the model optimises for what looks statistically plausible rather than what is structurally correct. Results have improved a lot as models got better at global consistency, and these remain the places errors show first.
Is diffusion related to how ChatGPT works?: They share the transformer machinery and the idea of learning a distribution, and they differ in the loop. A language model predicts the next token in a sequence. A diffusion model refines a whole image over repeated passes. Some newer systems blur that line by using diffusion for text or transformers inside the denoiser.

Related explainers

Multimodal AI Video and world models Embeddings

Part of See How AI Works, a free interactive course, where you learn how modern AI works by operating it, not watching videos.