How to Write an AI Image Prompt That Gets What You Pictured
Most disappointing AI images are not the model's fault. They come from prompts that describe a feeling rather than a picture — and a diffusion model has no way to render a feeling. It renders nouns, light and medium.
What follows works on Stable Diffusion 3 Medium, which is what ImagiForge runs, and on essentially every other diffusion model. None of it is model-specific trickery; it is just a description of what these systems are actually able to act on.
Five things worth describing
In roughly this order. You do not need all five — but a prompt with subject, light and medium will beat a paragraph without them almost every time.
Subject
The one thing the image is about, stated plainly. "A fisherman" beats "a person who works on a boat" — diffusion models resolve concrete nouns far better than descriptions of them.
an elderly fisherman
Setting
Where the subject is, and what surrounds it. This is what stops the model from defaulting to a grey studio backdrop, which is where it goes when you give it nothing.
on a weathered wooden dock at dawn
Light
The single highest-leverage word in most prompts. Light determines mood more than any adjective you could add, and models respond to it precisely.
low golden sunlight, long shadows
Medium
Photograph, oil painting, 3D render, screen print. Without one the model averages across all of them, which is why untethered prompts often look vaguely digital and unplaceable.
35mm photograph
Detail
One or two specifics that could only belong to this image. Specificity is what separates a picture from a stock image — but stop at two, because a long list dilutes every item in it.
salt-stained yellow oilskins
The five, assembled
Stack the examples above in order and you get a prompt that is specific without being long — about twenty words, every one of them doing work:
an elderly fisherman on a weathered wooden dock at dawn, low golden sunlight, long shadows, 35mm photograph, salt-stained yellow oilskins
Compare it to “a beautiful realistic portrait of an old man, highly detailed, masterpiece, 8k”. The second is longer and says almost nothing: no setting, no light, no medium, and five words spent on quality tags the model already applies.
Five habits that make images worse
Stacking quality words
"Masterpiece, best quality, 8k, ultra detailed, award winning" adds nothing a modern model does not already do. These tags are a habit inherited from older checkpoints and they mostly waste prompt budget that a real detail could have used.
Writing a paragraph
Attention thins out across a long prompt. Past roughly forty words the later clauses start being ignored, so the ending you cared about is the part that quietly gets dropped.
Negating in the prompt
"No cars" tends to produce cars — the model sees the word and weights it. Say what you do want instead, and leave exclusions to the negative prompt, which ImagiForge applies for you.
Asking for text in the image
Diffusion models render letterforms as decoration rather than language. Short words sometimes survive; a sentence will not. Add real text afterwards in an editor.
Describing two subjects at once
Attributes bleed between subjects — ask for a red car and a blue house and you may get a blue car. One subject per image is the reliable path.
Change one thing at a time
Diffusion is not deterministic, so two runs of the same prompt give different images. That makes it tempting to rewrite everything when a result disappoints — and it is exactly the wrong move, because you lose the information about which change helped.
Change the light. Or the medium. Or one detail. Keep the rest fixed. You will learn what each lever actually does far faster than by rewriting the whole prompt each time — which matters here, since you have a limited number of generations a day.