How to Write Prompts for Midjourney (Image Style)

Prompting an image model is not like prompting a chatbot. You are not writing a sentence, you are listing ingredients. Midjourney reads a prompt as a weighted bag of concepts, so what you say first and how you phrase each modifier both matter. The reliable structure is subject, then style, then parameters.

Start with a concrete subject

Lead with the thing you actually want to see, described concretely. A vague subject produces a generic image. Add the details that define it: who or what, doing what, where, in what light.

a weathered lighthouse keeper repairing a brass lamp,
inside a stone tower at dawn, warm light through a salt-stained window

Layer the style

After the subject, name the visual treatment: medium, art movement, lens, mood, color palette. Keep these as short comma-separated phrases rather than full sentences. Each phrase nudges the image toward a look.

Set the parameters

Parameters go at the end, each as a flag. The two you will use most are aspect ratio and stylize. --ar sets the shape of the canvas, and --stylize controls how much artistic liberty the model takes.

a weathered lighthouse keeper repairing a brass lamp,
stone tower at dawn, 35mm photo, golden hour, shallow depth of field
--ar 3:2 --stylize 250

• --ar 16:9 wide, for landscapes and banners.
• --ar 2:3 tall portrait, for characters and posters.
• --ar 1:1 square, for icons and product shots.
• --stylize lower values follow the prompt literally; higher values lean into the model's own taste.

Weight and exclude with intent

When two concepts compete, you can bias one with :: weights, for example forest::2 city::1 to favor the forest. To push something out of the frame, use --no, as in --no text, watermark.

The image prompt recipe

1. Subject: the concrete thing, described with a few defining details.
2. Style: medium, art style, lighting and camera as short phrases.
3. Parameters: aspect ratio and stylize, plus any weights or excludes.
4. Iterate: change one variable per run so you know what moved the result.

The fastest way to improve is to treat each render as an experiment. Keep a prompt that nearly works, change one phrase or one flag, and compare. That loop teaches you the model's vocabulary far faster than reading a list of magic words.