Midjourney Repeat: --repeat, Permutations and Batch Jobs (2026)

TL;DR: --repeat (or --r) re-runs one prompt several times from a single submission. Midjourney gates the count by plan: Basic 2 to 4, Standard 2 to 10, Pro and Mega 2 to 40. It runs in Fast and Turbo only, never Relax, and every repeat spends Fast GPU time as a separate prompt would.

What does --repeat actually do in Midjourney?

It presses the button again for you. Nothing more.

The Repeat article opens with the reason the parameter exists: "Every time you submit a prompt to Midjourney, you'll get a unique image—even if you use the same prompt." So --repeat is not quality control or a refinement pass. It is a shortcut for typing the same line several times: "you can tell Midjourney to run your prompt multiple times. For example, if you want three sets of images from the same prompt, just add --r 3 to the end of your prompt."

Two operational details matter. First, repeat leaves no trace: "the repeat parameter will be stripped out of your finished images, so you won't see it once your images are done generating. This also means if you re-run the prompt you will have to manually add the repeat parameter again." Every other parameter you can read back off a finished image. This one you cannot, so your notes are the only record a set came from one batch.

Second, Discord makes you confirm: "the bot will ask you to confirm the number of repeats and you'll need to click Yes before your prompts will run." The web Imagine bar has no such gate, which makes an accidental --r 40 an expensive typo.

a cast-iron teapot on a slate counter, window light --r 4 --v 8.2

One more thing before you plan a comparison. In the Legacy Features chart the Repeat row is ticked under "Affects Initial Images" and blank under "Affects Variations & Remix". Repeat multiplies fresh generations, not variations of something you already like.

Which versions and plan tiers is --repeat documented for?

The article says V8.2. The compatibility chart says nothing at all, and that gap is the interesting part.

The Repeat article carries the standard compatibility banner, worded "This feature is supported in V8.2", and V8.2 is current: the Version article states that "The current default Midjourney version is V8.2", which "released as the default version on July 24, 2026". Permutations carries an identical banner.

But the Feature Compatibility chart in that same article, where every careful parameter answer normally comes from, has no Repeat row and no Permutations row. Thirty-six rows covering aspect ratio, chaos, seed, tile, stylize, quality and multi-prompting, and neither batching feature appears. On the current model line, each article's own banner is your only published support statement.

A separate chart, in the Legacy Features article, does have a Repeat row: it ticks V6, V5, V4, V3, Test and Testp, Niji 4 and Niji 5. Niji 6 and Niji 7 appear in neither chart, so repeat support on the current anime line is undocumented. Our Midjourney version flags reference tracks which line is current.

Plan gating is where the real constraint lives, and there Midjourney is unambiguous: "The Basic plan allows 2 to 4 repeats, Standard allows 2 to 10, and Pro and Mega plans offer 2 to 40."

The Comparing Midjourney Plans chart states the same limit as a row called Maximum Repeat / Permutation Size. Note that it is one shared row, not two.

A footnote resolves the obvious worry: "Queue size will extend to accommodate repeat/permutation jobs." A 40-job permutation need not squeeze through a 10-job queue.

How do --repeat and permutations interact with Fast, Relax and Turbo?

They are locked out of Relax entirely, the most consequential fact on this page.

The Repeat article: "the repeat parameter only works in Fast and Turbo Modes. It won't work in Relax Mode." Permutations says the same, and the GPU Speed article states it from the other direction, listing under Limitations that "Permutation prompts, the repeat parameter, HD resolution videos, and Max Upscale (a legacy upscaler) are not available while using Relax mode."

Read that against what Relax is for. The GPU Speed article describes Relax time as unlimited, so on Standard and above you can generate all month without touching your Fast allowance. Batching is excluded from precisely the mode that costs nothing.

Turbo is more awkward than it looks. The Feature Compatibility chart marks the Turbo Mode row unsupported in the merged V8.1 and V8.2 column, while the GPU Speed article's body says "You can enjoy Turbo Mode with Midjourney versions 5 and later" and a banner atop that same article says "Turbo mode is not currently supported in V8.1." Three statements, one vendor, unreconciled at the time of writing. The article does promise a fallback: "If the Turbo GPUs aren’t available or you're on an older version, your job will switch automatically to Fast Mode."

So the practical answer for batching on the current version is Fast, which suits comparison work anyway. Turbo carries two penalties: it "uses double the Fast time for each image", and the Seeds article says outright, "Don't use Turbo mode when using seeds. Turbo mode focuses on speed, and as a result seed locking is not reliable." A seed-locked sweep in Turbo has the control removed. For the full speed-mode picture, see Midjourney's quality parameter and speed modes.

What does a batch actually cost you?

Midjourney's GPU cost table prices a standard-resolution prompt at 0.8 minutes and an HD prompt at 1.3 minutes. Its plan chart gives Basic 3.3 hr/month (200 minutes) of Fast time, Standard 15 hours, Pro 30 and Mega 60. Neither page joins them up, so the arithmetic below is ours, derived from their figures rather than published.

At 0.8 minutes a job, Basic's 200 minutes buys roughly 250 standard-resolution jobs a month. A single --r 4 is four of them, so one four-repeat batch a day spends about half the month. On Pro, 1,800 minutes makes a full 40-job permutation cost around 32 minutes, under two percent of the allowance. The gap between plans is not really the job cap. It is how much you can afford to be wrong.

Two things make that worse. Fast time does not roll over: the GPU Speed article is explicit that "any unused time doesn't carry over to the next month". And HD costs more, at 1.3 minutes against 0.8, so an HD batch is roughly 60% dearer for the same job count, which argues for sweeping in standard resolution and saving HD for the winner. Extra time has a list price too: the plan chart puts additional GPU time at $4/hr, so thirty jobs of exploration is about 24 minutes, or roughly $1.60.

How do permutation prompts work, and how many jobs do they make?

Braces, commas, and simple multiplication.

The Permutations article tells you to put your options in curly braces separated by commas. Its base case, a {red, green, yellow} bird, becomes three prompts, one per colour. Two brace sets multiply rather than pair up: of a {red, green} bird in the {jungle, desert} it says "will produce four prompts". Braces nest too, and the nested A {bird {on a pier, on a beach}, dog {on a sofa, in a truck}} also "will produce four prompts".

Crucially for sweep work, braces are not restricted to prompt text: "You can use permutations on any part of your prompt, including parameters," and the article's own example is a ratio sweep, --ar {1:1, 2:3, 3:5}. A backslash before a comma stops it acting as a separator, so {red, pastel \, yellow} bird yields two prompts rather than three.

The cap is the plan cap, per prompt: Basic tops out at 4 prompts from one permutation, Standard 10, Pro and Mega 40, the same row as the repeat limit. Multiply your brace sets out before you submit, because a 3 × 3 × 2 prompt is 18 jobs and a 4 × 4 × 3 prompt is 48.

The method: change one thing, and make the comparison mean something

Batching is only useful if the results are comparable. Four rules, in order of impact.

Isolate one variable per sweep. A permutation that changes the lighting and the lens and the stylize value produces images differing in ways you cannot attribute. You pick a favourite and learn nothing transferable. Change one axis; keep every other word and parameter identical.

Lock the seed when comparing, not when exploring. Midjourney's Seeds article makes the case in its own words: lock a seed "when testing different elements in your prompts" because "It’s like having a control in an experiment." Seeds run 0 to 4294967295, and two documented limits apply. The Version chart annotates the V8.x seed row "(99% identical)", so a locked seed is a strong control rather than a perfect one, and the Weird article notes that "Weird is also not fully compatible with seeds", ruling out a seed-locked --weird sweep. Our guide to --seed covers seed retrieval on web and Discord.

Use permutations for deliberate axes and --repeat for the variance floor. A permutation asks which option is best. A repeat asks how much the prompt varies when nothing changes, which is the baseline you need before claiming any option beat any other. Running --r 4 on your control prompt first is the cheapest sanity check there is, and it often shows the difference you were chasing is smaller than the model's own run-to-run spread.

Write down what you expect before you look. Anchoring is the failure mode of batch review, and it is not a discipline problem: the first acceptable image in a grid of 40 becomes the reference point for the other 39.

One wrinkle if you batch with random style references. The Style Reference article says that "If you use --sref random with a permutation, repeat, or V8.1 Draft mode prompt, each image will have a different style code." A random style reference and a controlled sweep are mutually exclusive.

Thirty-three copy-paste batch prompts

The model flag is --v 8.2, the current default. Swap it for whatever is current when you read this, and check the --ar reference before permuting ratios, since decimals are rejected.

Single-variable sweeps. One axis, everything else frozen.

studio portrait of a ceramicist at her wheel, soft key light --s {0, 100, 250, 1000} --seed 4041 --v 8.2

brutalist concrete library interior, wide shot --c {0, 15, 35, 60} --seed 4041 --v 8.2

single ripe fig on white marble, overhead --ar {1:1, 4:5, 3:2, 16:9} --seed 4041 --v 8.2

lone lighthouse in fog, painterly --weird {0, 250, 800, 2000} --v 8.2

a chef plating scallops, {rim light, soft window light, hard overhead light, candlelight} --seed 4041 --v 8.2

vintage motorcycle in a garage, shot on {24mm, 50mm, 85mm, 135mm} lens --seed 4041 --v 8.2

an armchair in an empty gallery, made of {boucle wool, cracked leather, moulded plywood, brushed steel} --seed 4041 --v 8.2

coastal cottage exterior, {dawn, midday, golden hour, blue hour} --seed 4041 --v 8.2

Permutation grids. Two axes, multiplied. Count the jobs before you send.

a {fox, heron} in a {snowbound orchard, flooded marsh}, naturalist illustration --v 8.2

product shot of an amber glass bottle on {white seamless, wet slate, raw linen}, {front-on, three-quarter} --v 8.2

a {tea house, record shop, barber shop} storefront at night, {neon signage, warm tungsten glow} --ar 3:2 --v 8.2

{watercolour, gouache, risograph} poster of a mountain range, {two-colour, four-colour} palette --v 8.2

A {cyclist {on a wet street, on a ridge line}, runner {in a stairwell, on a beach}}, documentary photography --v 8.2

a {crimson, pale \, dusty pink} peony, macro, dark ground --v 8.2

minimalist logo mark for a coffee roaster, {geometric, hand-drawn, monoline} --s {50, 250} --v 8.2

Seed-locked comparisons. The seed is the control. Keep it identical across the set, and out of Turbo.

a wooden sailing dinghy at anchor, calm water --seed 118822 --s 100 --v 8.2

a wooden sailing dinghy at anchor, calm water --seed 118822 --s 700 --v 8.2

a wooden sailing dinghy at anchor, {calm water, choppy water} --seed 118822 --v 8.2

an empty diner booth at 3am --seed 900001 --raw --v 8.2

an empty diner booth at 3am --seed 900001 --v 8.2

a rain-soaked city crossing --seed 771203 --no umbrellas --v 8.2

a rain-soaked city crossing --seed 771203 --v 8.2

still life with pears and a copper pan --seed 505050 --ar {1:1, 4:5} --v 8.2

Repeat prompts. Measuring variance, and spending the allowance on purpose.

a heron taking off from still water, backlit --r 2 --v 8.2

a heron taking off from still water, backlit --seed 660410 --r 4 --v 8.2

hand-lettered menu board in a bakery window --r 4 --fast --v 8.2

concept sketch of a folding electric cargo bike, marker rendering --r 10 --v 8.2

abstract cover art for a jazz record --sref random --r 6 --v 8.2

Two need a note. The --r 4 with a fixed seed is the variance floor test above, and Midjourney's "(99% identical)" annotation is what it measures. The --sref random --r 6 has the opposite intent: random style codes resolve per image inside a repeat, so you get six styles rather than six takes on one. Discovery, not comparison.

Triage prompts. These go to a language model, not to Midjourney, once the batch is in front of you.

I generated 12 images from one prompt with only the lighting term changed:
[paste the prompt and the four lighting values]

For each image I describe below, tell me which lighting term it most likely came from
and what specifically gives it away. Then tell me which visual differences are
attributable to lighting and which are probably run-to-run variance.

Image descriptions: [paste your notes]

Here is my creative brief: [paste brief]
Here are short descriptions of 8 candidate images: [paste descriptions]

Score each candidate 1-5 against each brief requirement in a table, with one sentence
of justification per cell. Do not name an overall winner. Flag any requirement where
all 8 candidates score 3 or below, because that means my prompt is missing something,
not that I picked badly.

This prompt produced the result I want: [paste winning prompt]

Rewrite it as a reusable template. Mark every part that should become a variable in
{{double braces}}, keep the parameters fixed, and list which parameters I should treat
as volatile because they depend on the current model version.

I ran a sweep of {list the values you tested} and the two best results came from
{value A} and {value B}. Design the next sweep: propose 4 values that explore the
range between and around A and B, explain what each one tests, and tell me what
result would prove the axis is exhausted.

Write a one-paragraph log entry for this batch: the prompt, the axis I varied, the
seed, the number of jobs, what I concluded, and what I would change next time.
Keep it under 90 words and make it searchable by writing the axis name explicitly.

That last one exists because --repeat is stripped from finished images: if you do not log the batch, nothing on the image says it was one. Save the family as a prompt template with the volatile parts marked, so the only edit when the default model moves is the version flag.

How do you review 40 images without anchoring on the first good one?

Change the order of operations. Judge against the brief before you judge against each other.

Write the pass criteria before you open the results. Three or four concrete requirements: the subject reads at thumbnail size, the hands are correct, the palette stays in range, the text is legible. Written afterwards, they describe whichever image you already liked.

Do a rejection pass first, not a selection pass. Go through the set once marking only fails. It is faster, less susceptible to order effects, and it usually shrinks 40 candidates to 8 before you make a single positive judgement.

Review once in reversed order. If the same images survive when you start from the end, order effects were not driving your shortlist.

Separate the best image from the best prompt. The strongest single result can come from the weakest option on your axis, through luck. What a metered batch buys is knowledge about the axis: if one option produced three of your eight survivors and another produced one, that is the finding, and it carries forward.

Then sort in Midjourney rather than a folder on your desktop. The Organizing Your Creations article documents filters, a timeline bar, Saved Searches, and a search over your own images covering "prompt text, as well as specific attributes, like parameters, individual reference images, Personalization profiles". The Using Folders article notes that generating while inside a folder puts results straight into it.

If a whole batch fails rather than splitting, the fault is upstream of the batching. Our image prompt troubleshooting guide works through the commonest causes.

What Midjourney does not publish about batching

State these as unknown, because the vendor does. Anything presented elsewhere as a rule here is somebody's inference.

• Whether --repeat and permutations can be combined in one prompt. No article says they can; none says they cannot. The single shared plan row hints at one job budget rather than two, but a hint is not a statement.
• What happens when you request more jobs than your plan allows. No documented error text, no documented clamping.
• Whether repeat jobs share a seed or each draw their own. The Seeds article is silent on repeat. Only the --sref random case is documented, and there each image differs.
• Repeat and permutation support on Niji 6 and Niji 7. No row in either chart.
• Whether batch jobs run in parallel, and how they meet the concurrent-prompt limits in the plan chart.
• Any GPU cost specific to batching. The cost table prices jobs, not batches. No discount, no surcharge.
• How --repeat interacts with --bs, the video batch-size parameter, which appears only in the Video and Parameter List articles.

How to check any of this yourself

docs.midjourney.com returns HTTP 403 to a direct fetch, which is why so much circulating Midjourney guidance is second-hand and years stale. The help centre runs on Zendesk and its JSON API answers normally: /api/v2/help_center/en-us/articles.json?per_page=100&page=1 enumerates every article, and /api/v2/help_center/en-us/articles/<id>.json returns one, each with an updated_at timestamp. That is how you date a claim rather than trusting a page's tone.

The IDs used above: Repeat 32757107922061, Permutations 32761322355597, Version 32199405667853, GPU Speed 32016412137741, Comparing Midjourney Plans 27870484040333, Seeds 32604356340877. Read the Version chart from raw markup or the icon cells look blank.

Then check the two numbers that decide affordability against your own account rather than any article, this one included: remaining Fast time on the Manage Subscription page or /info in Discord, and your plan's job cap.

Prompt Architects builds the prompt, not the image. If you want a Midjourney prompt with its parameters already correct before you spend GPU time on a sweep, image prompt generation starts on our Pro plan, and the Free plan includes 5 prompt enhancements per day, forever, per our FAQ. The generating still happens in Midjourney, on Midjourney's meter.