Write prompts that beat the default AI look

Sign in and confirm your model

Go to midjourney.com and sign in with a Google or Discord account. Midjourney now runs primarily in the web app, not Discord.

Midjourney requires a paid subscription, so plan to run economically: expect eight to twelve deliberate generations for this lesson, not fifty. In settings, confirm you are on the current default model (V8.1 was promoted to default on June 11, 2026); V7 is still selectable if a brief needs it.

Split one real brief into subject and style

Take an actual brief from your current work, not a toy prompt. Write it as one line, then split it into two buckets: subject (what is literally in frame) and style (how it looks: lighting, color, mood, medium).

Your text prompt will carry the subject; a reference will carry the style. Keep style adjectives out of the text if you plan to attach a reference, because the two signals fight each other.

Generate a baseline with explicit parameters

In the Create bar, write your subject-first prompt and append parameters at the very end, each marked with a double hyphen. Set --ar for your aspect ratio first, since changing it later forces you to redo the look.

Add --stylize (default 100; drop toward 0 for literal product work, raise toward 1000 for more artistic license) and --no to exclude unwanted elements like text or watermarks. Generate one set as your baseline.

Attach a style reference to kill the generic look

Two routes. Route A: add --sref random, generate, then save the numeric code Midjourney returns by liking it so it lands in your Likes tab.

Route B: click the image icon in the Imagine bar, upload your own reference at 1024px or larger, and drag it specifically into the Style Reference section, not the Image Prompt zone. The Style Reference pulls only the look; the Image Prompt would also drag in composition and subject.

Tune one variable per run

Re-run the exact same subject prompt and change only one thing each time. First adjust --sw (style weight, 0 to 1000, default 100) to control how hard the reference pushes.

Then adjust --stylize. Add --sv explicitly so the reference algorithm is pinned and does not drift later.

Changing one variable at a time is the whole point: it is the only way to know what actually moved the result.

Cap references at two and lock the winner

If you blend references, use two at most and give each an explicit weight; averaging three or more produces a muddy midpoint of nothing. When one combination clearly lands, stop iterating.

Note the exact winning prompt and every parameter value before you lose track of which run it was.

Write the versioned spec and prove it on a second subject

Record the spec in a doc: exact subject phrasing, the --sref code or reference image, --sw, --stylize, --ar, --sv, and model version. Label it v1.

Then swap only the subject text, keep the entire style block identical, and generate again. If the look carries cleanly to a new subject, the spec is real and reusable rather than one lucky seed.