You type "a cozy, warm café" into an image generator. You get a pretty picture, but a bland one that looks like a thousand others. You change three words, everything changes, and you have no idea why.
It's not a lack of talent. It's a lack of method. Here's the one we use on Ovrelia's Prism island: the five-layer grid.
The basic principle: describe visual facts
An image generator starts from random noise and turns it, step by step, into an image that matches your description. It's very good with what can be seen, and poor with what can only be felt.
If you could photograph it, the model can generate it. If it's an emotion, translate it into objects and light.
"Warm" can't be photographed. "An amber lamp, a steaming mug, rain on the window, late afternoon" photographs just fine.
The five layers
1. The subject
Who or what, in one precise sentence. "A bookseller in her sixties, short gray hair, shelving books." Not "a woman."
2. The setting
Where and when. "In a small neighborhood bookshop full of dark wood, on a winter evening."
3. The style
The look. Realistic photo, watercolor, flat illustration, 3D render, 1960s poster, engraving. One style at a time, or the model blends them.
4. The framing
Think like a photographer: wide shot, medium shot, close-up, high angle, low angle, front view. Framing completely changes the story the image tells.
5. The lighting
The most underrated layer. Golden dawn light, colorful neon, backlighting, soft window light, candlelight. Lighting creates the mood you were trying to get with your adjectives.
An example, layer by layer
- Subject: a bookseller in her sixties shelving books
- + Setting: in a small dark-wood bookshop, on a winter evening
- + Style: realistic photography
- + Framing: medium shot, slightly from the side
- + Lighting: warm light from a desk lamp, bluish shop window behind
The assembled prompt:
A bookseller in her sixties with short gray hair shelves books in a small dark-wood bookshop on a winter evening. Realistic photography, medium shot slightly from the side, warm light from a desk lamp, bluish shop window in the background.
The rule that changes everything: one change at a time
When you don't like the image, don't rewrite everything. Change just one layer, and write down what it changed in a small log:
| Try | What I changed | Result |
|---|---|---|
| 1 | Base prompt | Too dark |
| 2 | Lighting: "desk lamp" → "late afternoon, low sun" | Good mood, blurry face |
| 3 | Framing: "medium shot" → "close-up" | Keeper |
Three lines in your log, and you know exactly what each word changed. Next time, you start straight from the right place.
Weak spots to check
Even the best generators still have weaknesses: hands, text in the image, and logos. Always zoom in before using an image. For text (a poster title, a slogan), it's often safer to add it afterward in a layout tool.
Before you publish
One question is enough: could someone believe the image is real? If so, say it was generated by AI. And read the tool's terms of use before any commercial use: they vary from one generator to another.
Going further
On Ovrelia, Prism island goes further: keeping the same character from one image to the next, making a twenty-second clip, publishing by the rules. Each chapter ends with a project, such as a poster for a fictional event.
The first chapter, "From noise to image," is free: play it now.
If you're just starting out, begin with the step-by-step guide to generate an image with AI: pick a tool, make a first try, review the result. The grid is also a starting point to make a video with AI, shot by shot. And to test your reflexes, start the friend challenge: ten timed questions about AI, and a score to beat.
Frequently asked questions
In what order should I write the five layers?
Start with the subject, then the setting, style, framing, and lighting. The order mostly helps you not forget anything: what matters is that each layer describes visible facts.
Why does the same prompt give two different images?
Generation starts from different random noise on every try. So two images from the same prompt can vary, and that's normal.
Should I write image prompts in English?
English works well with every tool. If you write in another language and a detail gets misread, try rephrasing more simply before switching languages.
How do I fix an image that's almost right?
Change one layer at a time and compare. If you change everything at once, you won't know what improved or ruined the result.