How do Midjourney parameters like --ar and --stylize work?

When you type /imagine in Midjourney's Discord server, everything you write becomes the prompt: a description of the picture you want. Parameters are the exception. Anything that starts with two dashes, such as --ar 16:9 or --s 1250, is read as a setting, not as part of the description. Midjourney's documentation calls them switches, and they go at the very end of the command.

A typical full command looks like this:

/imagine prompt: a lighthouse on a rocky coast at dusk, watercolor --ar 3:2 --s 1250

The words describe the scene. The two switches say "make it wider than it is tall" and "go easy on the house style". With default settings, the bot returns a grid of four images in roughly 50 seconds.

The parameters worth knowing first

The current algorithm, version 3, has been the default since late July. These are the switches its guide covers that most beginners use:

• --ar (or --aspect) sets the aspect ratio, the proportion of width to height. --ar 16:9 gives a wide frame of about 448 by 256 pixels before upscaling. It replaces setting width and height separately with --w and --h, which the documentation says are unstable above 512.
• --stylize (or --s) sets how strongly Midjourney applies its own sense of style. The default is 2500. The user manual describes 625 as nearly switching styling off, 1250 as less strict but still pretty, 20000 as starting to drift away from your text, and 60000 as possibly ignoring the prompt altogether.
• --seed picks the random starting noise the image grows from. It takes a whole number between 0 and 4294967295. Without it, Midjourney picks one at random.
• --version (or --v) chooses the algorithm. Version 3 needs no switch; --v 2 brings back the previous one, and --v 1 the original, which the guide says can suit textures and close-up shots.
• --quality (or --q) sets how much rendering time to spend. The default is 1. A value of 0.5 is about twice as fast and costs half as much; 2 takes twice as long and costs double, for more detail.
• --no is a negative prompt: --no grass tries to keep grass out of the picture.

One prompt, six variations

The table below starts from one base prompt and changes one switch per row. The "what to expect" column follows the documentation's descriptions; your own images will differ, because each run starts from new random noise unless you fix the seed.

Base prompt: a lighthouse on a rocky coast at dusk, watercolor

Two practical notes. First, switches combine freely, so rows 3 to 6 keep --ar 16:9 to make them comparable. Second, --q 0.5 is a cheap way to run a table like this while exploring, since it halves the cost of each grid; switch back to the default once you know which settings you want.

How do you get consistent results?

Consistency comes from holding everything constant except the one thing you want to change. The seed is the main tool. The user manual says a fixed seed can sometimes make results steadier and more reproducible, which is a careful choice of words: it narrows the variation without promising identical pictures. To find the seed of an image you already like, react to the job's message with the envelope emoji, and the bot sends you its seed number and job ID by direct message. You can then reuse that number while you tweak the wording.

By default, each of the four images in a grid uses a different slice of the seed, which is why they vary. Adding --sameseed applies the same seed across the whole grid, so the four images change only slightly from one another. That is useful when you want to judge a small change in wording rather than browse for ideas.

A few habits do the rest. Keep the version fixed, because a different algorithm changes the whole look. Keep the aspect ratio and stylize value the same across a series. Change one word or one switch at a time, and write down what you changed. Finally, save your usual settings: the /settings command stores default parameters, and /prefer option set lets you create a personal shortcut, so a single custom switch can stand for a whole set such as a fixed ratio and stylize value. Then every image in a project starts from the same recipe, and only the subject changes.