How AI makes images

Priya types “a glass of cutting chai on a rainy Mumbai window ledge, steam rising” into an image tool. A few seconds later, there it is: the chipped glass, the grey sky, drops on the window. That exact picture never existed before.

It was not copied from the internet. It was not painted stroke by stroke from the top-left corner either. It began as fuzzy static, the kind an old TV showed when the antenna slipped. This lesson shows how static turns into chai.

At each step, the model looks at the noisy image and asks one question: “Which part of this looks like noise that should not be here?” It removes some of that noise. Then it looks again. Shapes appear first, then colours, then fine details.

Did you notice? At step 0 it is pure noise. Around the middle, a rough cup shape appears, though you could not swear it was chai. Steam lines show up around step 7, and by step 10 it is a clear pixel cup of chai on a saucer. Real tools work the same way, with far more pixels and usually more steps.

Classic diffusion models use roughly 20 to 50 steps. Newer fast models, trained to take bigger strides, can produce a good image in a handful of steps, sometimes fewer than ten. More steps usually means more detail, and more waiting.

Here is the clever trick. Making noise takes a computer no effort at all. So during training, we take a real photo with its caption, say “a cat on a scooter”, and add a little noise. Then a bit more. Then more, until it is pure static.

At every level of noise, the model is shown the noisy picture and the caption and must guess the noise that was added. We know the right answer, because we added it. So we can check the guess and nudge the model, the same guess, check, nudge loop you met in module 1.

After seeing hundreds of millions of picture and caption pairs, the model has learned what noise looks like on top of cats, chai, cricket grounds and Kathakali dancers. Then, when we give it pure noise and a caption, it “cleans” its way to a picture that fits the words.

Your prompt is first turned into embeddings, the lists of numbers that capture meaning, from the embeddings lesson. These numbers are fed into every denoising step. So at each step, the model removes noise in a way that moves the picture closer to “chai, rainy window, steam”.

Cleaning millions of pixels many times is slow. So most modern tools do the denoising on a compact sketch of the image, in latent space. A separate part, the decoder, then paints that sketch out into a full-size picture at the end.

This explains something you may have seen. Type the same prompt twice and you get two different pictures. The words were the same, but the starting static was different. Many tools also have a guidance setting: higher means “follow my words more strictly”, lower gives the model more freedom.

AI video tools stretch the same idea across time. Instead of cleaning one picture, they clean a whole short clip of frames together, so the cup stays the same cup as the steam moves. That is much heavier work, which is why video takes longer and costs more.

• Exact details can still slip: a long line of small text, the right number of objects, or “the red cup to the left of the blue one”. Models have improved a lot here, but check before you use the image.
• Pictures reflect their training data. Ask for “a doctor” or “a CEO” and you may get the same kind of face again and again. Be specific about who you want to see.
• Who owns the training pictures, and the style of living artists, is still argued in courts and parliaments in 2026. Rules differ by country and change often.
• Fake photos of real people can cause real harm. Many tools refuse them, and some add invisible watermarks or “content credentials” that show a picture was made by AI.