Midjourney Prompts and Parameters: From Luck to Control

By the end of this guide you will be able to run Midjourney as a repeatable process instead of a slot machine: you will know how to order a prompt, what each parameter actually changes, when to use draft mode, and what the subscription is really paying for.

Why most first attempts fail

The typical first prompt is something like "a girl in a cafe". It names a subject and nothing else: no lens, no light, no medium. The model fills the gaps with the statistical average of its training data, which is exactly why the result looks generically AI: centred composition, even lighting, soft detail.

Three things decide quality. First, how much information you supply about scene, light and material. Second, whether you constrain the output with an aspect ratio, a style and a model version. Third, whether you leave the model room to explore. That last one is counterintuitive: a one-line prompt produces four near-identical images, while a detailed prompt plus a little --chaos produces four genuinely different directions.

The seven-slot prompt

A prompt you can reproduce has seven fixed slots, filled in order: subject, action or state, environment, style and medium, lighting, composition and lens, parameters. Fixing the slots means that when you change one variable you can tell which variable did the work, instead of rewriting a paragraph each time and guessing.

Here is a complete example:

a ceramic coffee cup on a wet stone table, steam rising, quiet cafe by the window in early morning, editorial product photography, soft side light from the left, 50mm lens, shallow depth of field, low angle --ar 4:5 --style raw --v 7

• Subject: one ceramic cup. Keep one hero per image and no more than two supporting elements. This is the precondition for nothing going missing.
• State: steam rising. A state word makes the frame feel like something is happening.
• Environment: a quiet cafe window in the early morning. This sets the mood and the background complexity.
• Style and medium: editorial product photography. Be specific rather than aspirational; "premium feel" constrains nothing.
• Lighting: soft side light from the left. This single slot separates a snapshot from a photograph.
• Composition and lens: 50mm, shallow depth of field, low angle. Lens language dictates perspective and compression.
• Parameters: ratio, version and style strength go at the end.

Midjourney accepts prompts in other languages, but V7 responds most reliably to English, especially for lens, lighting and material vocabulary. A good habit is to think in your own language, then translate into short English phrases rather than long literal sentences. Then edit down: words like "beautiful" or "stunning" add no constraint and only dilute the real instructions.

Parameters that actually change the image

Parameters sit at the end of the prompt. Order does not matter, but keep them out of the descriptive sentence.

• --ar 16:9 sets the aspect ratio. Common choices are 16:9 landscape, 9:16 vertical, 4:5 for social posts, 1:1 for square. Decide the destination before you decide the ratio.
• --v 7 pins the model version. V7 is noticeably better at hands, in-image text, multi-element instruction following, realistic portraits and skin texture.
• --style raw reduces the default beautification. The output follows the prompt more literally and looks more photographic. Useful for product and documentary work.
• --stylize 0-1000 defaults to 100. Higher values push artistry and drift from the prompt. Lower it when fidelity matters.
• --chaos 0-100 increases variation across the four outputs. Use it while brainstorming, then bring it back down for final frames.
• --no text, watermark, extra fingers is an exclusion list. Negating items here works better than pleading for their absence in the body of the prompt.
• --sref <image URL> locks a visual style through a reference image.

One capability people underuse is character or object reference, which keeps a face or a product consistent across images. For a series, it is far more reliable than describing the same person over and over in words.

Draft mode: test composition cheaply

Draft mode renders faster and costs less, which makes it the right tool for the composition stage. The sensible sequence is: explore three or four compositions in draft mode, pick one, render it properly, then consider upscaling and local edits.

Most people do the opposite. They polish the wording first, generate four images they do not like, rewrite, generate again, and burn an hour. When the composition is wrong, better adjectives cannot save it. Draft mode is not mainly about saving credits; it compresses the decision into five minutes so you can see four directions at once.

Consistency across a set

The hard part of a series is not one good frame, it is making ten frames look like the work of one photographer. Three ingredients are enough: a style reference, a character or object reference, and fixed parameters.

1. Produce one master image whose composition, light and style all work.
2. Archive its prompt and settings verbatim, including ratio, version and stylize value.
3. Change only subject and action afterwards. Leave the style, lighting and lens sentences untouched.

Style drift almost always comes from rewriting the prompt for every frame. Your wording shifts each time, so the model's interpretation shifts with it.

Three common failures and fixes

• Missing elements: too many subjects. Put the most important element in the first fifteen words, cut supporting elements to two, and exclude the rest with --no.
• Broken hands and text: avoid complex hand poses such as overlapping fingers, gripping or gesturing. When text must be exact, stop fighting the generator and set the type afterwards. V7 handles in-image text far better than earlier versions, but it is still not a layout tool.
• Wrong aspect ratio: ask where the image will live before generating. A wide banner needs the subject left of centre with clean space on the right for a headline. Generating a square and cropping later usually decapitates the subject.

What you are really paying for

As of November 2026, monthly plans are Basic at $10, Standard at $30, Pro at $60 and Mega at $120, with annual billing around twenty percent cheaper, roughly $8, $24, $48 and $96 per month. There is no free tier: the free trial was withdrawn in 2023, so the entry point is $10. Prices and allowances are subject to change; check the official site.

The tiers differ mainly in GPU time. Basic includes about 200 minutes of fast time and no unlimited slow mode. Standard adds unlimited Relax generation plus 15 hours of fast time. Pro adds stealth mode, so generations stay private, and 30 hours of fast time.

Paid plans include commercial use from Basic upward. For context, images made during the old free trial could not be used commercially, though that tier no longer exists. Most people should start on Basic to validate the need, then move to Standard once they generate daily, because unlimited Relax mode is what makes batch composition testing affordable.

Practical constraints outside the US and EU

Being direct about it: using the service from mainland China requires your own network access and an overseas payment method, and this article does not walk through workarounds. If those are hard to arrange, running the same workflow on a domestic tool is the more realistic path. Prompt structure and parameter logic transfer; what changes is the interface and the quota rules.

A good next step: take one real brief, write a seven-slot prompt, explore four compositions in draft mode, pick one and finish it. One full pass teaches you more than ten lists of tips. For the wider pipeline, continue with the AI image workflow guide; for tool selection, see the AI image tools introduction; for side-by-side options, browse the tool directory.