Midjourney Prompts

How Midjourney reads a prompt

Midjourney isn't a chatbot. A standard prompt is a description of the picture you want to see, not a set of instructions, and its own guide says short, clear prompts usually work best. «Make me a picture of a lighthouse, make it stormy and paint it in watercolor» is a request. «Watercolor painting of a lighthouse in a storm at dusk» is a picture.

A good prompt decides five things and leaves the rest to the model:

1. The subject, with a number. «Three cats» or «a flock of birds» beats «cats», which can mean any count.
2. The medium. Photograph, watercolor, screen print, 3D render, film still. Often it's a single word, and it moves the result more than any adjective.
3. The setting and the light. Where it is, the time of day, and which side the light comes from, since a low sun from the left and flat overcast light give you two different pictures of the same scene.
4. Color and mood. A short palette and one feeling: muted oat and rust, quiet, playful.
5. The framing. Close-up, eye level, overhead, wide establishing view, with a lens if it's a photo.

Then the parameters, which always go at the end, after a space, with no commas or periods inside them. A pattern you can type from memory:

[Medium] of [subject] in [setting], [light], palette of [colors],
[framing] --ar [shape] --s [number] --raw --no text, watermark

Two habits from the docs save a lot of rerolls. Say what you want, not what you don't: a party «with no cake» can still grow a cake, so exclusions go in --no. And words you want inside the image go in double quotation marks, kept short; single quotes don't work.

The parameters that matter, as of October 2026

The current default model is V8.2, released on July 24, 2026. These are the parameters worth knowing in it, with ranges and defaults from Midjourney's documentation:

One spelling to know about. Older guides write Raw as --style raw; Midjourney's current docs spell it --raw, and that's what the blocks on this page use.

Older tutorials also use --cref (Character Reference) and --oref (Omni Reference) to put a character or object into a picture. In V8.x both are replaced by the Edit model, which takes up to four reference images at once. And if an old style code from a tutorial looks wrong, the docs explain why: style references changed between V6 and V7, and adding --sv 4 brings back the older behavior.

Which prompt for which picture

Each block below is one kind of image with the parameters already set. Pick by the job, then fill the brackets.

One prompt, rewritten

This is the sort of prompt that still circulates from older tutorials:

a beautiful cozy cabin in the woods, ultra realistic, 8k, highly detailed,
masterpiece, award winning, trending on artstation --v 6

Here's how we'd write the same idea today. It's our example, written for this page, not something we're presenting as model output:

Photograph of a small timber cabin in a snowy spruce forest at blue
hour, warm light in two windows, smoke rising from the chimney, one
set of footprints leading to the door, shot on 35mm from the path at
eye level, palette of deep blue and amber --ar 3:2 --s 150 --raw
--no people, text, watermark

Half the original describes nothing you could see. «8k», «masterpiece» and «trending» name no light, place or lens. The rewrite spends the same words on things the picture can show: the hour, two lit windows, smoke, a single set of footprints that suggests someone just arrived. «--v 6» goes too, because it pins an old model; keep it only if you want V6's look. Raw and a modest stylize keep it reading as a photograph, and the --no list stops a stray hiker or a watermark from turning up.

Fixes for a grid that missed

Change one thing at a time, so you can tell what helped.

Where Midjourney prompts go wrong

Failed prompts tend to break in the same places.

• Instructions instead of a description. «Make it brighter» or «copy this style» reads poorly in a standard prompt. Describe the brighter picture. The exception is the Edit model in V8.x: with an image attached, it does take instructions like «show this from behind».
• Negatives in the sentence. «Without trees» puts trees in the model's head. Use --no.
• Broken parameter syntax. The docs list the usual mistakes: no space before the dashes, a space between them, a comma after the value, or more description after the parameters.
• Too many style words with a reference. With --sref, the docs advise keeping the text simple and describing content, since style words can pull against the reference.
• Real people. Our blocks only describe invented people. Attach a photo of a person to the Edit model only if you have the right to use it.

Plans, website and Discord

Every block pastes as written into the Imagine bar on the website or into Discord. Reference images go through the image panel on the web and as URLs in Discord. As of October 2026, Midjourney's help center lists four plans at $10, $30, $60 and $120 a month, with 20% off when you pay for a year. Standard and up add unlimited image generation in Relax mode, and only Pro and Mega include Stealth mode, which keeps your images private. There's no free trial on the website or in Discord. Our Midjourney review keeps the full plan table, the AI image generators page compares the alternatives, and the AI image prompts page has blocks written for tools that read plain sentences.