A prompt is a description, not a command. The model reads it as a sentence and paints what it understands, so the fastest route to a better image is a more concrete sentence — not a longer one.
What actually happens between the sentence you type and the image that comes back.
An AI image generator reads your prompt as a description, not as a command. It has learned which arrangements of pixels tend to go with which words, and it builds a new picture that matches the description you gave. Nothing is retrieved from a library and…
Subject, setting, light, framing and finish — the checklist behind every good result.
Most prompts that disappoint have the same shape: a subject and nothing else. A prompt that works usually carries five things, and you can write them as one plain sentence rather than a list of keywords.
What the style buttons really do, and why writing the look into your prompt is sometimes better.
Every style in the picker maps to a fixed phrase that is placed in front of your prompt. Photorealistic sends "A photorealistic image of", Watercolour sends "A watercolour painting of", and so on for the whole list. That is the whole mechanism — there is no…
Why cropping afterwards ruins a composition, and what 1K and 2K actually change.
Aspect ratio is not a crop applied at the end. It is sent to the generator as a parameter before the image exists, so the composition is built for that shape: a 16:9 result places the subject for a wide frame, and a 9:16 result rebuilds the same idea for a…
When to upload an image, what the brush does, and why the two paths give different results.
Generating and editing answer two different questions. Generating answers "make me a picture of this". Editing answers "keep this picture, but change one thing about it". If your goal contains the words "this photo", you want the editing path — upload the…
Negatives, keyword soup, contradictions and the other habits that quietly cost you images.
The first mistake is writing what you do not want. "No people" mentions people, and mentioning a thing makes it more likely, not less. Describe the empty street instead — "an empty street at dawn" gets you what "a street with no people" does not.
One or two sentences is usually enough. What helps is being concrete about the subject, the light and the framing, not adding more words: a prompt that reads like a sentence beats a pile of comma-separated keywords.
Why does the same prompt give a different image every time?
Every result is generated from scratch rather than retrieved, so the same sentence run twice gives two different pictures. To keep a composition and change one detail, upload the result and edit it instead of generating again.
Do negative prompts work?
There is no separate negative field, and writing “no people” tends to make people more likely rather than less, because mentioning a thing brings it in. Describe what should be there instead: “an empty street at dawn”.
Does picking a style change the model?
No. A style adds a fixed phrase to the front of your prompt before the request is sent, and that exact phrase is printed under the prompt box. Same model, same credit.
Should I pick the aspect ratio before or after generating?
Before. The ratio is sent as a parameter, so the composition is built for that shape; cropping a square into a banner afterwards throws away half the picture.