Write image prompts that get the composition you meant
Subject, framing, light and style are four separate decisions. Most prompts only make one.
Before you start
- Access to any image generator
- An image you can picture but cannot describe yet
What you will be able to do
- Write prompts as four explicit decisions rather than one description
- Control framing and light instead of accepting the default
- Iterate on one variable at a time so you learn what did the work
A vague prompt gets you a competent, generic image — centred subject, even light, mid-shot, nothing wrong with it and nothing you asked for.
Image prompts are not sentences so much as a short spec, and the parts people leave out are always the same ones.
Name the subject and only the subject first
Get the thing itself right before you touch anything else.
Start with the plainest possible statement of what is in the frame, and generate. It will be generic — that is fine, you are checking that the model understands the subject at all.
If the subject is wrong here, no amount of style language will rescue it, and people routinely spend twenty prompts on lighting for an image whose subject was never right.
Add framing, because the model will not guess it
Shot size, angle and where the subject sits in the frame.
Three separate things: how close (close-up, mid-shot, wide), from where (eye level, low angle, overhead), and where in the frame (centred, off to one side, filling it).
Left unsaid, all three default to the middle of the distribution, which is why unspecified prompts look interchangeable. Borrowing the vocabulary of photography and film works well here because the training data is labelled with it.
Set the light explicitly
Light does more for the mood than any adjective you can add.
Direction, quality and time of day: "low sun from behind", "soft overcast", "single hard light from the left". These change the image far more than words like "dramatic" or "moody", which the model has to interpret into lighting anyway — badly, and differently each time.
- If the result feels flat, it is almost always the light rather than the subject or the style.
Change one thing per iteration
Rewriting the whole prompt teaches you nothing about which word worked.
When something is nearly right, resist the urge to rewrite. Change a single element, regenerate, and keep the version that improved.
This is slower per attempt and much faster to a result you can reproduce — and after a dozen images you will have a genuine sense of which words this particular model responds to, which does not transfer from reading prompt guides.
- Stacking five style keywords at once. When it improves you will not know which one did it, and half of them are probably fighting each other.
Subject, framing, light, style — name all four and the hit rate changes immediately. Then change one at a time.
Common questions
Was this guide useful?
100% of readers found this useful
Read next
Upscale and retouch AI images without the plastic look
The waxy, over-smooth quality people associate with AI images is mostly introduced after generation, by upscalers doing exactly wh…
Get consistent characters and style in AI image generation
Generating one good image is easy. Generating twelve that clearly show the same person in the same style is the actual problem — a…
Connect two apps with an AI step in the middle
The genuinely useful automations are not the clever ones. They are a trigger, one AI step that makes a small judgement, and a writ…
Choose your first AI assistant without overthinking it
Every comparison table lists twenty differences and only three of them change your day. Here is how to pick in ten minutes and get…
