What --sref actually does, measured across 55 images

--sref is the most useful parameter in Midjourney and the least explained. The documentation says it applies a style reference. What it does not say is what a bare number actually refers to, how reliable it is, or what else has to be true for a code to reproduce the look you saw someone else get.

This collection is a reasonable place to answer that, because every image in it ships with its real prompt and the parameters are stored as structured data rather than as a string. So the questions are just queries.

• 55of 85 images carry a --sref
• 31distinct style codes
• 11codes used more than once
• 7images on the single most-used code

A code is an address, not a style

The mental model that trips people up is thinking of --sref 3808932062 as naming a style the way art deco names a style. It does not. The number is closer to a seed: it addresses a fixed point in the model's style space that Midjourney can return to. It carries no meaning, it is not searchable, and there is no index of them — which is exactly why collections like this one exist.

Two consequences follow, and both show up in the data. First, a code is only as good as the record of it: lose the number and the look is gone. Second, a code constrains style but not subject — the same code across different prompts gives you different pictures that feel related.

The eleven codes that repeat

Twenty of the thirty-one codes here appear exactly once, which tells you nothing. The eleven that repeat are where the signal is:

Notice that most codes stay inside one category. That is the "style, not subject" claim showing up as a measurement rather than an assertion: a code that produces a soft film look keeps producing a soft film look, so the images it makes keep landing in the same bucket.

Seven images, one code

The most-used code here is --sref 3808932062, on seven images. They are not seven versions of one picture — the subjects are a child on a bicycle, a man walking under autumn trees, a beach, a portrait. What they share is the treatment:

Every image in the collection using --sref 3808932062. Different subjects, one recognisable film-stock look.

This is what a working style code buys you: you stop describing the rendering and start describing the subject. Compare the prompt for the first of them — there is almost nothing in it about how the image should look, because the code is carrying that:

Everyday snapshot, shot on Svema color film --ar 3:2 --v 7 --raw --sref 3808932062 --p n59emtm 72csxc4 --s 50

The pattern: codes travel with a stylize value

Here is the finding that is not in any documentation, and it fell out of cross-referencing two parameters that are usually discussed separately.

Eight images in this collection use --stylize 50. Seven of those eight are the seven images above — the entire --sref 3808932062 group. Ten images use --stylize 250, and eight of those ten fall into exactly two style-code groups: --sref 536544871 and --sref 4114080079. The two images at --stylize 240 are the only two carrying --sref 1742458012.

The causal story is straightforward once you see it. --stylize controls how much of Midjourney's own aesthetic opinion gets layered on top of your instructions. A style code is an instruction about style. Turn stylize up and the model's defaults start overriding the code you asked for; turn it down and the code comes through clean. So a code that was found at a low stylize value only reproduces at a low stylize value.

Two codes at once

--sref takes more than one value. Two pairs in this collection are used that way — 117928474 2167081065 on two images, and 1742458012 3361981308 on two more. Midjourney blends them.

The four images built from two style codes at once. The first pair share 117928474 2167081065; the second pair share 1742458012 3361981308.

Worth knowing when you go looking for codes: a multi-value --sref is two addresses, not one compound address. Each number works on its own, and this site treats them separately for exactly that reason — --sref 117928474 and --sref 2167081065 each get their own page.

What a code will not do for you

• It is not a guarantee across versions. Style codes address a point in a particular model's space. The images here were made across --v 6.1, --v 7 and --niji 6; a code recorded under one is not promised to mean the same thing under another.
• It does not fix a weak description. The seven images above still say what is in them. The code decided the treatment, not the subject.
• It is not the same as a personalisation profile. Twenty-four images here use --profile instead, which points at an account's learned preferences rather than a shared coordinate — and unlike a style code, someone else's profile is not something you can simply paste.

Where to start

Pick one of the repeating codes above, open its page, and read the four or seven prompts that use it together. Seeing one code against several different subjects tells you what it holds constant far faster than running it yourself. Then copy the code and the stylize value into a prompt of your own.

The full set of style-code pages is linked from the collection, and every image page lists the exact parameters behind it.