One Code. Infinite Variations. How Midjourney's --sref Locks Your Visual DNA.

Here’s the problem with finding an image you love in Midjourney: you can’t reproduce that exact aesthetic on a different subject.

You found a set of icons generated with a specific illustrative quality — the way shadows sit under forms, the specific hardness of edges, the warmth of the color temperature. You try to describe it in words. “Flat illustration, warm palette, soft shadows.” You get close. Not there.

--sref (style reference) is the solution. Pass the URL of an image, or a Midjourney-issued sref code, and the model extracts the visual grammar of that image — color relationships, edge treatment, texture, compositional biases — and applies it to whatever you generate next.

It’s not copying. It’s vocabulary transfer.

Full prompt (copy this exactly):

What each part is doing

vintage Japanese woodblock travel poster composition

This is a style-category term, not a description of what the image should look like. It tells the model to bias toward the compositional conventions of that specific genre: the layered distance recession, the high-horizon approach, the large flat foreground silhouette. Without this anchoring, even a strong --sref can produce something that looks woodblock-influenced but doesn’t hit the genre conventions.

three layered mountain ridges in descending tonal values

You’re describing the structural logic of the image, not just the content. “Three ridges” gives the model a specific count. “Descending tonal values” tells it how depth is represented — lighter in the distance, darker in the foreground. This aligns with how woodblock prints actually handle depth, which reinforces the --sref signal.

washi paper texture background

Material specificity. The model has training data for washi paper as a distinct material category — it’s not “textured background” (vague), it’s a specific paper with a specific visual signature. Matching your material vocabulary to the --sref’s visual world means both signals point the same direction.

limited four-color palette

A constraint, not a description. “Limited” forces the model away from its default tendency toward rich tonal variation. “Four-color” gives a specific ceiling. In combination with the --sref, this helps preserve the flat, graphic quality of the source aesthetic instead of accidentally adding photographic complexity.

--sref 2847163950 --sw 80

The style reference code and style weight. The code points to the source aesthetic; --sw 80 sets how strongly it’s applied. --sw 100 means style can completely override subject description — sometimes useful, usually too aggressive. 80 is the working sweet spot: strong enough to enforce the aesthetic, permissive enough to let your subject description still land.

3 variations to test

Variation 1: Swap the subject, keep everything else

Same code, different scene. The test: do they look like they were made by the same hand? If yes, your sref is working. If the fishing village looks completely different in character, lower --sw and verify you’re not fighting it with too-specific subject description.

Variation 2: Reduce style weight for subject freedom

At --sw 50, the sref is a recommendation, not a mandate. Modern subjects fight traditional aesthetics when the style weight is too high. Drop the weight, let the subject breathe.

Variation 3: Same sref, completely different category

Product shot using a travel poster aesthetic. This is the brand-building use case: generate catalogs, marketing images, content series — all sharing the same visual DNA even across radically different subjects.

Model compatibility

Midjourney v6.1: Full --sref support. The current implementation is stable and consistently reproducible across days if you use the numeric code (not a URL that might change). This is where --sref was designed to live.

DALL-E 3: No --sref equivalent. The closest proxy is a detailed style description: “in the visual style of [specific aesthetic descriptor].” Less precise, but DALL-E 3 responds well to specific art movement and material references.

Stable Diffusion XL: IP-Adapter achieves similar results. Run your reference image through IP-Adapter at 0.6–0.8 influence — same logic as --sw, different interface. ControlNet style transfer is an alternative but harder to get consistent.

Gemini Imagen 3: Style description in the prompt only. The model is strong at following art-direction language but has no URL-reference mechanism. Use precise category terms and material words rather than general “aesthetic” language.

When this breaks

1. Subject-style genre conflict. A hypermodern architectural render doesn’t want to live inside a woodblock grammar. The model splits the difference and you get something that looks confused. Fix: increase or decrease --sw to let one signal dominate, or pick a sref that’s genre-compatible with your subject.

2. Vague source sref. If the sref you’re using comes from an image that’s “editorial photograph” or “digital painting” — broad categories — the style extraction will be too generic to do useful work. Better srefs come from images with distinctive visual identities: a specific illustrator’s work, a defined printing technique, a period-specific poster aesthetic.

Run this prompt and compare outputs at --sw 40 and --sw 100. The difference between those outputs will tell you more about how Midjourney’s style reference system actually works than any explanation could. The model at 40 treats the sref like a mood board; at 100, it’s a mandate. Learn where the useful range lives for your aesthetic.

Share your sref codes in the comments. Codes that get shared publicly are the ones that end up in everyone’s toolkit. That’s how the best visual languages in this community stop being locked to one person’s prompt history and start becoming shared vocabulary.