Images
AI can make a picture from a few words, or change a photo. It's like fog that slowly clears, and your words decide what appears. Learn how to describe what you want, and when an AI image is the wrong choice.
First, what do you think?
You ask an AI for "a picture of my grandma at the beach". What will it look like? Have a guess.
How does AI make an image?
The whole story in three sentences. An image model starts with a fuzzy mess of dots. Step by step, it makes the mess a little sharper. Your words decide what appears.
Think of thick fog. Slowly the fog clears. Bit by bit you see a shape, then a house, then the windows. With AI, your description decides what appears in the fog.
Researchers call this a diffusion model. It learned to undo noise: to turn a fuzzy image into a slightly sharper one.
- Making. From a description, it makes a new image. A dragon on a skateboard, a poster, an invitation.
- Editing. You give it a photo, and say what to change. Google lists adding, removing or changing parts, and changing the style.
- Labelled. Google puts a hidden watermark, called SynthID, in every image its model makes.
What the predictor does
The image model predicts what fits your words. That's the rule.
Ask for "my grandma at the beach". The model has never seen your grandma. It makes a grandma, one that fits the words. The more you describe, the closer it gets to what you have in mind. But it will never be her.
A short description leaves a lot open. Which beach? Sunny or windy? A photo or a drawing? The model fills the gaps with what's common. Google's own advice: describe the scene in rich detail. The more specific you are, the more control you have.
The fog clears into what fits your words. Describe more, get closer. But it's never a photo of something real.
Try it yourself
Three image tasks. Which description or step gives you more control?
Everything runs in your browser. Nothing you type is sent anywhere.
How do you get the image you want?
Four steps. Try them today on something fun, like a card for a friend.
Describe the scene
Say what's in it, where, and what's happening. Not "a dog". Better: a small brown dog on a red sofa, in the morning sun.
Name the style
A photo, a pencil drawing, a cartoon, a watercolour. Bold lines or soft colours. Google says: be specific about the style, for consistent results.
Improve in a conversation
Change one thing at a time. "Make the sofa blue. Don't change anything else." Google recommends improving images in a conversation like this.
Say it's made with AI
Share it as made with AI. In the EU, deepfakes must be labelled. Being open is also fair.
In practice
Three people make an image. The steps, and the limits, make the difference.
Case 1The party invitation
Lisa (12) wants an invitation with a dragon for her birthday. She types "dragon" and gets a scary monster.
Give up, or use the scary one.
"Dragon" leaves everything open. The model filled the gaps with what's common.
Describe it. A friendly green dragon with a party hat, blowing out candles. As a colourful cartoon. Then change one thing at a time.
Every detail gives the model more to aim for.
Case 2The bike for sale
Tom (52) sells his old bike online. His photo looks dull. He thinks about making an AI picture that looks shinier.
Use the nicer picture. Everyone does it.
The buyer expects to see this bike. An AI picture shows a bike that fits the words, not his bike.
Take a better real photo, in daylight. Use AI to write the advert text instead.
A real photo is honest. The buyer gets what they saw.
Case 3The joke about a classmate
Jesse (14) wants a funny picture of a classmate as a clown. He wants to share it in the class group.
Think it's harmless: everyone knows it's fake.
An image like that can look real. The Fraudehelpdesk calls a copy of someone's face a deepfake. It's hard to tell from the real thing. It can hurt, and it spreads.
Don't make images of real people without asking them. Make the clown a made-up character instead.
Nobody gets hurt, and the joke is still there.
When not to
- Not as proof of something realIt shows what fits, not what happened.Use a real photo for news, sales or evidence.
- Not of real people without permissionIt can look real, and hurt.Ask first, or make up a character.
- Not without saying soPeople may think it's real.Say it's made with AI.
By situation
AI images can brighten a poster or presentation. Say they're made with AI. For a report about something real, use real photos with a source.
Use AI images for ideas, sketches and decoration. Not in place of a real product photo. Check your employer's rules about AI images.
Cards, invitations, a picture for a story. Describe the scene and the style, and improve it step by step. Leave real people out, unless they agree.
Myths and mistakes
| Claim | True? | Why |
|---|---|---|
| It can draw my grandma | No | It never saw her. It makes a grandma that fits the words. |
| One word is enough | Not usually | It fills the gaps with what's common. Describe the scene and the style. |
| A fake image is always obvious | No | A deepfake can be hard to tell from the real thing. |
| Images of real people are fine as a joke | No | They can look real and hurt. Ask first. |
Related topics
Read first
- Text and chatthe same idea: describe, then improve.
Next
- Audio and videoAI that listens, speaks and makes video.
Related
- Deepfakes and scamswhen fake images are used to deceive.
Sources
- researchHo, Jain en Abbeel (2020). Denoising Diffusion Probabilistic Models.Diffusion models make images by learning to undo noise: they start from a fuzzy mess and make it a little sharper, step by step.
- documentationGoogle AI for Developers. Image generation with Gemini.Describe a scene in rich detail: the more specific, the more control. Be specific about the style. Improve an image in a conversation. You can add, remove or change parts, or change the style. All images it makes carry a SynthID watermark.
- governmentEuropean Commission. AI Act.The EU AI Act sorts AI into four risk levels. Some uses are banned, like social scoring. People must be told when they talk to a chatbot, and deepfakes must be labelled. AI literacy rules apply from 2 February 2025.
- organisationFraudehelpdesk. Fraudevormen en -termen.A deepfake copies someone's voice or face with AI and is hard to tell from the real thing. In 'help request fraud' a 'son' or 'daughter' asks for money by app, sometimes with a recorded copy of their voice.
Test
You get 7 questions about the whole page. Get 5 right, and you've mastered the topic.
About the tools in this course. Every tool is described in the same way, in alphabetical order, without a recommendation. Names, prices and features change often: the maker's own site is always right. This site was built with the help of Claude, one of the tools described here.