Midjourney Image Prompt Anatomy With Worked Examples

A Midjourney image prompt means two different things depending on who's searching. Midjourney uses the phrase for a specific feature, an image you attach to the prompt so the model borrows its content, composition, and colours. Most people typing it mean a prompt that makes a Midjourney image. This article covers both, because they're the same anatomy. A prompt has up to four inputs, text, image prompts, style references, and an identity reference, plus parameters at the end, and nearly every bad result people describe comes from using the wrong input for the job.

Text says what's in the frame. An image prompt lends the picture's content. A style reference lends its look and nothing else. An omni reference, or the edit model that replaces it in V8, puts a specific person or object in. Parameters set the job. Get those five sorted and the prompt writes itself, so the rest of this is the docs' own numbers for each input and one subject built six ways.

One honesty note before the numbers. Midjourney's documentation site refuses automated readers, so every fact below comes from archived copies of the docs pages, snapshotted between December 2025 and September 2026, with the dates in this article's ledger. Where a page was older than the current model, I say so.

Five Inputs, One Text Box

Here's the anatomy as a table, with the weight parameter for each input and the rules the docs attach to it.

The last row is thin on purpose. The parameter list describes --edit in one line, "create and modify images using written instructions and up to four reference images", and marks omni reference as "replaced by the Edit Model in V8.X". The edit model's own docs page had no archived copy I could open, so I'm not going to describe behaviour I haven't read. If you're on V8.2, which the version page lists as the default since July 24, 2026, the identity slot in your Imagine bar is labelled Edit Model Reference, and the omni rows above describe the V7 feature it replaced.

The image weight row is the one to reread. The default is 1 and the ceiling is 3, which is a narrower dial than the style and omni weights, and the docs add that "different Midjourney versions may have different image weight ranges". If a tutorial tells you --iw 5, it's describing a version that isn't the one you're on.

Parameter Grammar, From The Docs' Own Mistakes

Parameters are where prompts silently fail, because a malformed parameter doesn't error, it becomes prompt text. The parameter list page gives three rules and four broken examples, and the examples are worth more than the rules.

Two more from the aspect ratio page, since --ar is the parameter everyone uses. It can't contain decimals, so 1.39:1 is written --ar 139:100. And you can paste pixel dimensions and let it simplify, so --ar 1920:1080 becomes 16:9 on its own. The default is square, and the page calls extremely wide or tall ratios "experimental".

One Subject, Built Six Ways

Same picture each time. A small red fishing boat pulled up on a pebble beach under an overcast sky. Each block shows the Discord form, where images are URLs in the prompt. On the web Imagine bar you click the image icon and drop the upload into the matching slot instead, and the notes say which slot. These are built from the documented forms. I haven't rendered them for this article, and the site's rule is that I don't describe results I didn't get.

Text Only

Colour photograph of a small red fishing boat pulled up on a pebble beach, overcast sky, flat grey light, muted palette, quiet mood, wide shot from low on the beach --ar 3:2

Prompt Basics lists seven areas to consider, subject, medium, environment, lighting, colour, mood, and composition, and that prompt fills each once in that order. The page's own example of what not to do is "Show me a picture of lots of blooming California poppies, make them bright, vibrant orange, and draw them in an illustrated style with colored pencils", against "Colored pencil illustration of bright orange California poppies". Same information, a third of the words. Their advice is that short prompts let the default style fill the gaps, and "fewer details mean more variety, but you get less control".

Image Prompt Plus Text

https://your-host.example/boat-photo.jpg small red fishing boat pulled up on a pebble beach, overcast sky, flat grey light, wide shot from low on the beach --iw 2 --ar 3:2

On the web, the photo goes in the Image Prompt slot and the text stays as written. The URL goes first in Discord, separated from the text by a space, and the image has to be online already. Two docs rules shape the text. It "should include everything you want in the final image, rather than instructions for changing the reference image", so you describe the boat, you don't write "make this boat bigger". And you should "crop your image to match the aspect ratio of your final image", which is why --ar 3:2 appears here and the photo should be 3:2 before upload. --iw 2 is two thirds of the way up the dial.

Style Reference

small red fishing boat pulled up on a pebble beach, overcast sky --sref https://your-host.example/gouache-painting.jpg --sw 200 --ar 3:2

On the web this image is dragged into the Style Reference section rather than the Image Prompt one. The reference here is a gouache painting of something else entirely, and that's the point. A style reference "doesn't copy objects or people, just the overall style", so the painting's subject is irrelevant and its colours, medium, and texture come across. The docs' best-practice list says to keep the text simple and to focus on content, not instructions, with "the look of this image but a dog" as the bad example and "detailed portrait of a dog" as the good one. --sw 200 is double the default on a dial that runs to 1000. If you'd rather not upload anything, --sref also takes a numeric code from Midjourney's style library, and --sref random picks one for you, which the site's midjourney prompts post uses as a vocabulary exercise.

Omni Reference

Gouache painting of a small red fishing boat pulled up on a pebble beach, overcast sky, gouache painting --oref https://your-host.example/my-actual-boat.jpg --ow 200 --ar 3:2

Now the reference is a photo of a particular boat, dropped into the Omni Reference section on the web, and the job is to put that boat in the painting. Three things from the omni reference page explain the shape of the prompt. The style word appears at the start and the end because the page says that if you want a different style from your reference you should "mention your desired style at both the start and end of your prompt". The weight is 200, above the default of 100, because the page says a lower weight needs the physical characteristics reinforced in text, and keeping it below 400 avoids results the page calls unpredictable. And the prompt describes the scene rather than the reference, for the same reason as the image prompt.

Two costs to know. Adding an omni reference "will automatically run the prompt in V7", whatever your default, and the page says it uses 2x the GPU time of a regular V7 image. The page also warns that "intricate details like specific freckles or logos on clothing may not perfectly match". On V8 the same slot is the edit model, and I'd expect the prompt shape to survive even where the numbers don't.

All Three References Together

https://your-host.example/beach-photo.jpg small red fishing boat pulled up on a pebble beach, overcast sky --iw 1 --sref https://your-host.example/gouache-painting.jpg --sw 150 --oref https://your-host.example/my-actual-boat.jpg --ow 150 --ar 3:2

The docs say a style reference "can be combined with Character References/Omni References and Image Prompts", so this is legal. The beach photo sets the scene, the painting sets the look, the boat photo sets the boat. I've pulled every weight toward the default because three references compete for the same image, and the omni page makes exactly that point about stylize and omni weight fighting each other. My concern with this shape is that when it goes wrong you won't know which reference did it, so I'd build it up one reference at a time and only run the full version once the pairs behave.

Multi-Prompt With An Exclusion

pebble beach under an overcast sky::2 small red fishing boat:: --no people, footprints --ar 3:2

The double colon splits the prompt into ideas Midjourney weighs separately, with no space on its left and one space on its right. ::2 makes the beach twice as important as the boat. The --no list at the end is the documented way to exclude things, and Prompt Basics is blunt about the alternative, "if you mention a party with 'no cake,' a cake might still appear". Two quirks from the No page. Moderation reads every word after --no on its own, so --no modern clothing is read as "no modern" and "no clothing", and the page suggests naming the clothing you want instead. And --no is the same as a multi-prompt weight of -0.5, so vibrant tulip fields --no red equals vibrant tulip fields:: red::-0.5, with the constraint that all the weights have to add up to a positive number.

One hedge on this block. The multi-prompt page I could open was archived in August 2025 and lists compatibility through version 6.1. The version chart has a multi-prompting row, but its checkmarks are images that didn't survive extraction, so I can't tell you from the docs whether V7 or V8 honours ::. The --no parameter has its own page with a V8.2 note on it, so that half is current. How exclusion behaves across other generators is in negative prompts.

Image Prompt Versus Style Reference

The related searches ask for this comparison directly, and it's the decision that matters most. Both take an image. They take different things from it.

An image prompt borrows content. The docs say Midjourney "looks at your image's core elements", and the article title promises influence over "content, composition, and colors". Use it when the reference has the thing you want in the frame, a layout, a colour scheme, a scene.

A style reference borrows look. It "doesn't copy objects or people". Use it when the reference has the finish you want and the wrong subject, which is nearly always, because the pictures you admire are rarely of the thing you're making.

The failure mode is using an image prompt when you meant a style reference. You get the reference's subject creeping in, and you fight it with --iw 0.5 and negative prompts when the fix was the other slot. The reverse mistake is rarer and easier to spot, because a style reference simply won't give you the reference's content no matter how high you set --sw.

Neither is a copier. The image prompt page says references are "inspiration to guide new creations, not to copy them exactly", and points anyone who wants precise changes to their own image at the Editor. Which brings up the other half of the People Also Ask box.

Can Midjourney Edit An Existing Image

Yes, on the website, and the Editor page describes it as working on "both Midjourney images and your own personal ones". Your own images go in through the Edit tab, by upload or URL. The tools are an Erase brush that marks what to regenerate and a Restore brush that brings it back, Smart Select for building a mask from include and exclude points, Move and Resize with aspect ratio presets, and Layers for compositing another image in. Retexture regenerates the whole image "using a new style while keeping the structure and composition of the original". Each submitted edit returns four results. There's also a Suggest Prompt button that runs Describe on your image, which is the closest thing to a built-in prompt generator.

The caveat is on the page's first line as of the July 2026 snapshot. V8.2 images can be opened in the Editor, "but Editor functionality is currently using V6.1". So the edit is done by an older model than the one that made the picture, and the version page adds that inpainting or outpainting an HD image downscales the result to SD. If the edit is a small region, that's fine. If the edit is most of the image, I'd rerun the prompt with a reference instead.

How do you ask it to edit your picture, in words? The Editor walkthrough answers with its own example. Extend the canvas to make room, erase the region where the new thing goes, and describe what you want to see in the erased area, a crown on the cat's head in their case. Describe the result, not the operation. The same rule that runs through every reference type runs through the Editor too.

And a line from both the Editor and omni pages that people skip. "You must have the necessary rights to use the images you upload." Uploads and edits of external images are visible only to you, per the Editor page, but the rights question is yours either way.

Questions From The Results Page

What are good prompts for Midjourney? Short ones that fill the seven areas once each, in the docs' own words, "short and simple prompts typically generate the best images". The text-only block above is the shape. The longer templates on the results page, forty of them on one site, are fine as vocabulary and long by the vendor's standard.

How do I generate a Midjourney image? Type the text in the Imagine bar on the website, click the image icon to attach references into their slots, add parameters at the end, and submit. In Discord the same prompt has image URLs at the front and --sref or --oref URLs at the back.

Is there an image prompt generator, a free one, or a PDF? The Editor's Suggest Prompt runs Describe on an image you give it, which is the vendor's generator. A prompt built by a tool still needs the anatomy above to be worth running, and the site's art prompt generator post is about making that generator yourself. Everything on this page is free to use with a subscription that's already paying for the GPU time. A PDF goes stale the day a version ships, and the version page lists three shipping between March and July 2026.

What transfers to other generators? The text, mostly. I run image models locally on an M4 Pro, and none of the local models have slots that mean what these three mean. The general rules that survive the move are in ai image prompt.

What I'd Do First

Take one photo you like and run it through two prompts, once in the image prompt slot and once as --sref, with the same short text. Look at what each one kept.

That's the whole distinction in two generations, and once you've seen it you'll stop reaching for --iw to fix a problem that was in the other slot. Then read the parameter grammar table once more and check your last ten prompts against it. In my judgment at least one has a comma in it.