Midjourney Seed: Reproducing a Result with --seed (2026)

TL;DR: The Midjourney seed locks the starting noise, nothing else. Midjourney's own Seeds article files it under a heading reading "Not for Saving Styles" and points you at style references, omni references and personalization instead. Valid range is 0 to 4294967295, default random, and on V8.1 and V8.2 results are annotated "99% identical".

What does the Midjourney seed actually control?

One thing: the random noise the render starts from. Midjourney's Seeds article calls a seed "the initial spark that sets your Midjourney image in motion" and compares it to television static, then draws the boundary in a single clause. Seeds "only influence the initial layout of noise that begins the rendering process."

That sentence is the whole post. A seed is an input to the first step, not a description of the output, and carries no information about colour, medium, camera, subject or character.

Two more properties, same page. You set it with a whole number "between 0 and 4294967295", and if you do not set one, "Midjourney will use a new random one every time, giving you a variety of outcomes." That default is why the Repeat article states that "Every time you submit a prompt to Midjourney, you'll get a unique image—even if you use the same prompt."

a cast-iron kettle steaming on a wood stove, cold morning light --ar 3:2 --seed 1234

Midjourney's own recommended use is narrow, and it is not reproduction. Under a thumbs-up icon: "Lock a seed when testing different elements in your prompts." And then: "It’s like having a control in an experiment."

Does the same seed and the same prompt give you the same image?

Close, not identical, and only inside one session. Most pages about the Midjourney seed get this wrong, and Midjourney publishes the correction twice: a compatibility badge on the Seeds article reading "Seeds are 99% identical in V8.X.", and the Feature Compatibility chart in the Version article, where the Seed Parameter row carries a check icon in the V8.1 and V8.2 column with "(99% identical)" appended. Nowhere does it define what the missing one per cent covers.

Then there is the time dimension, which is worse. The Seeds article's "Not Always Predictable" section says "Seeds may behave unexpectedly, so they shouldn't be relied on for the same results over different prompting sessions", and follows it with unusually blunt advice: "Assume that if you take a break and come back later, your seed may no longer work as expected."

Which Midjourney versions accept --seed, and what is the valid range?

Every model line Midjourney documents, with the caveat attached only to the current one.

The range is confirmed three times: the Seeds article gives it in prose, and two default-value tables in Legacy Features list Seed at 0-4294967295, default random.

The Version article dates the current model line precisely: "V8.2 released as the default version on July 24, 2026." A seed tutorial written against V5 is describing a different model and a different chart. Our Midjourney version flags reference tracks which line is current.

Why does a seed stop working the moment you change anything?

Because it is the weakest input in the stack, and Midjourney says so in as many words. The "Not Always Consistent" section of the Seeds article states: "While seeds can help with consistency, they have the least impact on the final image." It then lists what outranks it: "Factors like your prompt, model version, and even slight parameter or settings changes can lead to unexpected results."

That is four failure modes.

You edited the prompt. The seed sets the noise; the prompt sets everything the model does with it. Change one adjective and the render diverges from the start. The seed is still doing its job. Its job is just very small.

You changed model version. A seed is an index into one model's noise handling. Swap --v 7 for --v 8.2 and you changed the function, not the input.

You changed a parameter or a setting. Note the word "slight". Toggling --raw, nudging --stylize, switching between SD and HD, or flipping a personalization toggle all count, and a setting you did not knowingly change is still a changed setting: the August 21, 2026 alpha changelog fixes three that changed themselves, including "Vary now respects your settings instead of quietly running in HD". Even a fixed --ar is not fixed end to end, since "Some aspect ratios may be slightly changed when upscaling"; our aspect ratio reference covers the caps per model line.

You used Turbo. The Seeds article has a section headed "Not for Use in Turbo Mode": "Don't use Turbo mode when using seeds. Turbo mode focuses on speed, and as a result seed locking is not reliable." That is close to moot on the current default, since the Version chart marks Turbo unsupported on V8.1 and V8.2 while the GPU Speed article contradicts itself, carrying a banner reading "Turbo mode is not currently supported in V8.1." above a body saying "You can enjoy Turbo Mode with Midjourney versions 5 and later." Our quality and speed reference unpicks that chart.

Does the seed carry through variations, Remix and upscales?

Nothing Midjourney publishes says it does, and two pages imply clearly that it does not. The Remix article lists which parameters survive: "Only parameters that normally influence variations will work while using Remix, including: aspect ratio, no, stop, and tile." Seed is not on that list.

The Legacy Features article's Parameter Version Compatibility chart is more explicit still. Two columns sit before the version columns, headed "Affects Initial Images" and "Affects Variations & Remix". The Seed row is ticked in the first and blank in the second, alongside chaos, image weight, quality and repeat, while aspect ratio, no, stop and tile are ticked, matching the Remix article's list exactly.

One documented case does carry a seed forward, and it is a Midjourney feature rather than a user technique. The Version article, describing the web button that reruns a standard-resolution job at higher resolution, says it "will rerun the seedlocked prompt in HD". The April 14, 2026 changelog introducing that button described it more loosely, as rerunning "any SD job as HD".

What does Midjourney recommend instead of seeds?

Three named tools, in the part of the documentation almost nobody quotes. Under a thumbs-down icon, the Seeds article carries a section headed "Not for Saving Styles":

"Seeds can’t capture or bookmark a specific style, character, or appearance across different prompts. They only influence the initial layout of noise that begins the rendering process. For consistency in your images, we recommend style references, omni references, and personalization."

That is the vendor calling the popular use case the wrong one, and naming three replacements.

Style references are the closest thing to a saved look. A reference "doesn't copy objects or people, just the overall style", and style codes are numeric handles into Midjourney's internal style library. Critically for reproducibility: "Once your prompt has a style code, using rerun/reroll or variations keeps the same style code." That is durability a seed does not have. One limitation before you plan around it: "You cannot create a style code based on an uploaded image." Our guide to --sref and style codes goes deeper.

Omni references pin a subject rather than a style. The Omni Reference article states that "Omni Reference is compatible with Midjourney version 7", and that "Adding an Omni Reference image to the Imagine bar will automatically run the prompt in V7." The Version chart agrees, marking it unsupported on V8.1 and V8.2 with the note "(Uses V7)". Character consistency currently means an older model line.

Personalization pins your taste, and the mechanic that matters is the code. A profile changes as you rate more images, and "As it updates, it will generate new codes that act like labels for the different versions." Copy the code and you have pinned a snapshot. Current state: "Your V7 Global Personalization Profile is compatible with V8.2", and "You can now create additional profiles for V8.X, however there is currently no Global V8 Profile."

Style codes are not immune either, which makes the point general rather than anti-seed. The Style Reference article: "The Style Reference feature was updated between V6 and V7, and older style codes may not produce the same styles anymore." The fix is a version pin, --sv 4. On V8.1 and V8.2 the --sv values and defaults are not documented at all: the only collapsibles on that page are titled "V7 Style Reference Versions" and "V6 Style Reference Versions".

How do you find the seed of an image you already made?

Yes, and on both surfaces, in the Seeds article itself.

On the web. Open any image from your Create or Organize page, use the options button near the prompt text, click Copy, then Seed.

In Discord. React to the finished job with the plain envelope emoji. The article is specific that it must be the plain one, and that the Midjourney Bot then sends the seed by direct message. If nothing arrives, check the Midjourney server's privacy settings and confirm Direct Messages is toggled on.

The neighbouring mechanism is the Job ID, covered in the Show Command article: same copy menu on the web, same envelope reaction in Discord, and /show <job_id> revives the job. It is the stronger handle, because it addresses a result where a seed addresses a starting condition.

One curiosity, in case an old tutorial hands it to you. --sameseed is listed under Legacy Parameters as a V3 feature: "When --sameseed is specified, all images in the initial grid use the same starting noise and will produce very similar generated images." The Legacy chart ticks it for V3 only. It is not a current parameter.

Is a seed ever a guarantee anywhere in AI?

No, and image generation is not the outlier. Text-model vendors document the same thing in incompatible ways.

Anthropic's Messages API reference has no seed field at all, and says of temperature that "the results will not be fully deterministic" even at 0.0. Google's prompting-strategies page takes the opposite line, stating that "A temperature of 0 is deterministic". OpenAI's OpenAPI specification does expose a seed, and documents it with the words "Determinism is not guaranteed". Three vendors, three positions, one shared property: nobody promises you the same output twice. We took that apart properly in temperature, seeds and determinism.

Midjourney sits comfortably in that company. "99% identical" is a hedge of the same species as "best effort", with the same practical consequence: build the workflow so that reproducing the asset matters more than reproducing the generation. Download the image. Save the Job ID. Save the full prompt string.

Twenty-one copy-paste prompts for seed testing and the alternatives

Swap the subject text, keep the structure. The first block is the one thing a seed is genuinely good for: an A/B with everything except one variable held still. Run both back to back in one session:

a lone fisherman on a stone jetty at dawn, muted palette --ar 3:2 --v 8.2 --seed 88123
a lone fisherman on a stone jetty at dusk, muted palette --ar 3:2 --v 8.2 --seed 88123

One parameter isolated against a fixed seed:

brutalist library interior, concrete and oak --ar 16:9 --seed 4021 --stylize 100
brutalist library interior, concrete and oak --ar 16:9 --seed 4021 --stylize 500
brutalist library interior, concrete and oak --ar 16:9 --seed 4021 --raw

Proving a seed does not cross model lines:

a paper crane on a windowsill, soft rim light --seed 777 --v 7
a paper crane on a windowsill, soft rim light --seed 777 --v 8.2

The two ends of the documented range, to check the parameter is parsed at all:

weathered brass compass on a nautical chart --seed 0
weathered brass compass on a nautical chart --seed 4294967295

A seed sweep to pick a starting point before you commit. The Permutations article states "permutations only work in Fast and Turbo Modes", and the Seeds article warns off Turbo, leaving Fast:

art deco travel poster of a mountain railway --ar 2:3 --fast --seed {101, 202, 303, 404}

Holding the seed while testing chaos, the deliberate opposite knob. Chaos is "set to 0" by default and ranges to 100:

overgrown greenhouse, humid air, shafts of light --seed 5150 --chaos 0
overgrown greenhouse, humid air, shafts of light --seed 5150 --chaos 45

Now the tools Midjourney recommends. A style code persists through reroll and variations in a way a seed does not, and random resolves into a code you can read off and reuse:

a chef plating a dish in a small kitchen --ar 4:5 --sref 1523098234
a chef plating a dish in a small kitchen --ar 4:5 --sref random
a chef plating a dish in a small kitchen, wider shot --ar 4:5 --sref 1523098234 --sw 250

Pinning the older style-reference model when a pre-V7 code no longer matches:

inked botanical study of a fig branch --v 7 --sref 2984771 --sv 4

An omni reference for a subject, remembering it runs the job in V7:

the woman from the reference standing at a bus stop in the rain, cinematic --oref https://your-image-url.png

Personalization, default profile then a pinned code:

a quiet fishing village at blue hour --ar 3:2 --p
a quiet fishing village at blue hour --ar 3:2 --p yourProfileCode

Style, personalization, seed, version and speed all pinned. As close to a repeatable Midjourney job as the documentation supports:

a ceramicist's hands centring clay on a wheel --ar 4:5 --v 8.2 --sref 1523098234 --sw 200 --p yourProfileCode --seed 88123 --relax

The same block as a reusable template, which is how it survives a real project:

[SUBJECT], [LIGHTING] --ar [RATIO] --v 8.2 --sref [STYLE_CODE] --sw [0-1000] --p [PROFILE_CODE] --seed [0-4294967295]

What Midjourney does not publish about seeds

Writing the gaps down beats filling them with plausible numbers, because every one is answered confidently by pages that invented the answer.

• What the missing one per cent of "99% identical" is. Not defined, not measured, not scoped.
• Whether --seed behaves differently on Niji 6 or Niji 7. Neither chart has a row for it.
• What happens to a seed across an HD and SD switch, beyond the rerun-button line.
• How long a seed stays valid. Being told to assume it stops after a break is a warning, not an interval.
• Whether an out-of-range seed errors or wraps. The range is published; the failure behaviour is not.
• --sv values and defaults on V8.1 and V8.2. Documented for V7 and V6 only.
• Any relationship between seed and style code. Two independent identifiers, never discussed together.

How to check any of this yourself

Midjourney's parameter surface moved repeatedly in 2026, so verification beats memorisation.

Read the compatibility chart, not the parameter article. The Feature Compatibility chart in the Version article is the only page stating per-version availability for seed, quality, Turbo, Draft and Omni Reference side by side. Where it disagrees with a topic article, compare update dates on both. On August 27, 2026 the Version article carried July 24 and the Seeds article carried July 27.

Test in one sitting, on one model, with one variable. That is the use Midjourney endorses, and the only condition where a seed behaves like a control.

Fetch the docs properly. docs.midjourney.com returns 403 to most automated fetches, which is not a dead link. The Zendesk Help Center JSON API behind it answers plainly, and updates.midjourney.com carries the dated changelogs. Every figure here came from one of those two on August 27, 2026.

The durable move is not memorising a seed number. It is keeping the whole prompt block, parameters included, as a reusable template with the volatile parts marked as variables, so that when the default model changes underneath you the only thing you edit is a version flag.