A picture out of noise
Lesson 1 of 13 in the course “AI Images and Slides”. Free, with animations and a quiz.
Meet Liya
Liya is 20, studies marketing and makes candles she sells online. She constantly needs pictures: product cards, posts, a poster for a craft fair. Plus a presentation for uni and decent photos without a photographer. There’s no money for a designer. Throughout the course we solve her tasks with AI, and you do the same for yours.
How AI draws a picture
Many models draw through diffusion. During training they saw millions of captioned pictures and learned to remove noise, the coloured static you’d see on an old TV. When you write a prompt, the model takes pure noise and removes it step by step, checking against your words. After a few dozen steps a candle emerges from the static. The models in ChatGPT work differently inside, but the point is the same: the model doesn’t search the web for a finished picture, it draws a new one to match your description.
Why it’s different every time
The noise at the start is new every time. It’s set by a random number called a seed. That’s why the same prompt gives four different candles. It’s a feature, not a bug: ask for 3–4 options at once and pick the best. And if one is almost perfect, don’t start over, edit it. That way you keep the good base and change only what you don’t like.
AI draws the “average”
The model learned what the thing you name usually looks like. Type “candle” and you get the most typical candle: white, in a glass, on a dark background. To get yours, describe what makes it different: the shape and colour of the wax, the label, where it stands, what the light is like. The rule is simple: the fewer words, the more choices the model makes for you, and the more the picture looks like thousands of others.
How fast it took off
In late March 2025 OpenAI built new image generation into ChatGPT, and feeds flooded with Studio Ghibli-style pictures. According to OpenAI, 130 million people made more than 700 million images in the first week. In late August 2025 Google released Nano Banana, and trends like “3D figurines from selfies” brought over 200 million photo edits and more than 10 million new Gemini users in nine days. AI images are no longer a geek toy, they’re an everyday work tool.
What’s next
Five levels, from easy to advanced. First you’ll see how AI draws and which service to pick, including free ones. Then you’ll learn to write prompts, add lettering and keep one style. Next come photos: edits in plain words, quality, a photoshoot, and why an AI passport photo gets rejected. Then presentations people actually listen to. At the end: rights, labels, fakes and a final project for your own thing.
Check yourself
- Liya sent the same prompt twice and got two different candles. Why?
Correct answer: Each time the model starts from new random noise. The starting noise comes from a random number (the seed). New noise, new picture. That’s why it’s handy to ask for several options and pick. - Liya typed “candle” and got a white candle in a glass, nothing like hers. What will help?
Correct answer: Describe her candle: shape, wax colour, label, background and light. The model draws the most typical thing. The differences of her candle have to be put into words, or the model has no way to know them. - Of four options, the third is almost perfect, only the background is wrong. What now?
Correct answer: Edit that option: ask to change only the background. It’s better to keep a good base and change one detail at a time. Starting over is a new lottery with new random noise.