Vague prompts get vague images. A simple six-part formula for AI image prompts — subject, action, setting, style, light and framing — with before-and-after examples.
Updated
Open all four studios free — no card, nothing to install.
If you've tried an AI image generator and got something that was almost right — the wrong mood, a weird angle, a background that makes no sense — the problem usually isn't the model. It's the prompt.
AI image models are very literal. They fill every gap you leave with the most average guess. "A coffee shop" gets you the most generic coffee shop imaginable. Describe the one you actually see in your head, and the result gets far closer.
A useful prompt answers six questions:
| Part | Question | Example |
|---|---|---|
| Subject | Who or what, exactly? | a young woman in a mustard kurta |
| Action | What are they doing? | laughing while holding a clay cup of chai |
| Setting | Where and when? | on a Jaipur rooftop at sunset |
| Style | What kind of image? | editorial lifestyle photograph |
| Lighting | What's the light like? | warm golden-hour light, soft shadows |
| Composition | How is it framed? | medium shot, eye level, 4:5 portrait |
Put together:
An editorial lifestyle photograph of a young woman in a mustard kurta, laughing while holding a clay cup of chai on a Jaipur rooftop at sunset, warm golden-hour light, soft shadows, medium shot at eye level, 4:5 portrait.
You don't need all six every time, but the more of them you decide, the less the model decides for you.
Weak: "a product photo of a candle"
Better: "A studio product photo of a handmade soy candle in an amber glass jar, on a travertine stone block, dried flowers beside it, soft window light from the left, beige background, minimal and calm, close-up, 1:1."
Weak: "YouTube thumbnail background for a tech video"
Better: "A dark desk setup with a glowing laptop and neon blue and purple rim light, lots of empty space on the right for text, cinematic, shallow depth of field, 16:9."
In both cases the second prompt says what's in the frame, how it's lit, what the mood is — and leaves room for the text you'll add later.
Style is the part people most often leave out, and it changes everything. A few useful families:
Add a mood word or two — calm, energetic, luxurious, playful, nostalgic — and the model adjusts colour and composition to match.
Tip: Avoid asking for a specific living artist's style or a real brand's look. Describe the qualities you like instead — "bold flat colours, thick outlines, playful characters".
These are the fastest way to make AI images look less "AI":
If the image is going on Instagram, ask for 4:5; for Stories and Reels, 9:16; for YouTube, 16:9. Generating square and cropping later often cuts off something important. (Not sure which size you need? See our Instagram size guide.)
Many image models still get lettering wrong — misspelled words, strange characters. For posters, thumbnails and ads, generate the image without text and ask for "empty space on the left" (or wherever your headline will go). Then add the text in a design tool, where you control the font and the spelling.
Your first result is a draft. To improve it:
If you change five things at once, you won't know which change helped. Generate a few variations each time; image models add randomness, so the best result is often the third or fourth.
In Vision Studio you can pick from several image models, set the aspect ratio before generating, and see the credit cost before you press generate. When you have the image you want, open it in Pixel Studio to add your headline, logo and colours — or turn it into a short video clip for a Reel.
Long enough to describe the subject, setting, style, light and framing — usually one to three sentences. Longer isn't better if the extra words are vague or contradictory.
Many image models still struggle with exact lettering. Generate the image without text, then add headlines in a design tool where you control the font and spelling.
It depends on the tool's terms and the content. On Editing Karo, paid plans include a commercial licence for what you generate — but avoid prompts that copy real people, brands or a specific living artist's work.
Image models add randomness to each generation. Generate a few variations and pick the best, then refine the prompt one detail at a time.
