Skip to main content
Every image or video task starts with the prompt field in your request body. This page covers how to write that string well. The guidance comes from upstream Midjourney documentation; which flags and values your prompt may use is defined by the image parameter matrix, which is generated from the backend validation spec. When the two disagree, the matrix wins.

​
Keep it short and specific

Short, descriptive phrases consistently beat long instructions. Describe the snapshot you want, not the process of making it.
Avoid: Show me a picture of lots of blooming California poppies, make them bright, vibrant orange, and draw them in an illustrated style with colored pencilsPrefer: Colored pencil illustration of bright orange California poppies
Three wording habits that measurably change results:
  • Use precise synonyms. “Big” is vague; “gigantic” or “enormous” gives the model a stronger signal.
  • Use numbers, not plurals. “Three cats” is controllable; “cats” is not. Collective nouns work too — “a flock of birds”.
  • Describe what you want, not what you don’t. Saying “no cake” in the text can still produce a cake. To exclude things, use the --no parameter instead of negations in the text itself.

​
The seven detail dimensions

A one-word prompt works — the model fills every gap with its default style. That means more variety and less control. Anything you care about must be spelled out. When a result disappoints, check which of these dimensions you left unspecified:

​
Vocabulary that moves the needle

Single adjectives in the right slot change the image dramatically. Useful starting points, grouped by what they control:
  • Artistic medium — ______ style cat: block print, ballpoint pen sketch, cyanotype, graffiti, risograph, ukiyo-e, watercolor, pixel art, cross stitch, oil painting, cut paper.
  • Time period — illustration of a ______ cat: 1700s, 1920s, 1950s, 1980s each carry a distinct visual language.
  • Emotion — ______ cat: shy, determined, joyful, sleepy.
  • Color — ______ colored cat: sepia, pastel, duotone, iridescent, grayscale, neon, acid green.
  • Environment — ______ cat: tundra, jungle, desert, salt flat, crystal forest, ocean.
Treat these as probes: swap one word at a time and compare, rather than rewriting the whole prompt between runs. For reproducible comparisons, pin --seed.

​
Anatomy of a full prompt

A prompt string can carry three kinds of content:
  1. Text — the description above. Required.
  2. Image references — image URLs that steer content or style, attached via image prompts or reference parameters (--sref, --oref, --cref). Which of these a given model version accepts varies sharply — --cref is rejected across the whole v8 family, and --oref exists only on v8.0. Check the parameter matrix before using them.
  3. Parameters — flags appended at the end of the string, e.g. --ar 16:9 --stylize 150. They control rendering, not content.
Pin a version explicitly. When --v is omitted the API renders as v7, and parameter support differs across versions — a prompt tuned on the default may fail or change meaning on v8.2. See Models for version selection.

​
Next steps

Once the basics feel predictable, move to Advanced Prompting: multi-prompts and weights, permutation prompts, negative prompting, and rendered text.