Add --sref to the end of your prompt, give it a style source, and Midjourney V8.1 renders your subject wearing that aesthetic — its palette, lighting, texture, and finish — while your prompt keeps the subject. The style source can be one or more image URLs, one or more numeric codes, or a blend of both. That is the whole mechanism. The craft is in which source you pick, how many you stack, and how hard you let the style push, and V8.1 changed enough defaults that older habits need a refresh.
V8.1 became Midjourney's default model on June 10, 2026. It renders roughly four to five times faster than earlier versions and ships HD — native 2048×2048 — by default, with no separate upscale step. Most relevant here: Midjourney describes srefs and moodboards in V8.1 as exceptionally stable, which is what lets a single style reference carry a whole series instead of quietly drifting subject to subject. This is the hands-on walkthrough; for the conceptual map of what a style code actually is, the complete sref guide is the pillar.
Paste one or more image URLs straight after --sref. Any direct image link works (a link ending in .png, .jpg, or .webp), but the most common source in practice is a Midjourney-hosted s.mj.run link — the short URL Midjourney mints for an image you generated or uploaded. These are personal and temporary, so they belong in your prompt history, not in a shared swatch book. Midjourney reads the image purely as a style guide: it borrows the look, never the subject.
Numeric srefs still work in V8.1. There are 232 of them — every integer from 0 to 4,294,967,295, about 4.3 billion distinct, repeatable aesthetics; drop one after --sref and you get the same style on any subject. One compatibility rule trips people up: numeric codes and --sref random only resolve under style versions --sv 4 (legacy) and --sv 6. The newer --sv 7 engine — now the default for image references, faster and HD-friendly — won't take a raw code at all. So if you feed a number, pin --sv 4 or --sv 6 or watch it get ignored.
The --sv 6 / --hd conflict. Running srefs or moodboards on --sv 6 costs about 4x the GPU time and does not work alongside --hd or --q 4. So you genuinely choose: the richer --sv 6 style engine, or native HD output — not both in the same job. For most code-driven work, --sv 4 keeps you HD-compatible and cheaper.
Image references behave differently from codes in one important way: an image is an interpretation of a specific picture, while a numeric code points at a fixed style vector. Codes are therefore more consistent across runs — which is exactly why people collect them. But image refs let you chase a look you already made and love, with no number to discover. Most real V8.1 work uses image refs.
Here is an actual portrait generated on V8.1 with image references. The prompt below is real — the two s.mj.run links are two of the references that fed it. Open the image to read its full decoded style DNA.
Note the caption: this image actually stacked nine references. We're showing two of them — proof that you are not limited to a single style source.
List several sources after one --sref, space-separated, and V8.1 blends them into a composite style. The real images behind this guide go deep on this — one in the set stacked twelve references into a single coherent look. You can mix freely: two image URLs, a URL plus a code, three codes. To bias the blend, append relative weights with the double colon, exactly like prompt weighting:
Stacking is how you invent a finish no single reference gives you. The discipline: many references pulling in compatible directions compound into richness; references fighting each other average into grey mud. The twelve-ref image works because every source shares a tonal world. Use weights to let one reference lead rather than letting five sources cancel out.
The real payoff of style references is series consistency. Both portraits below were generated from the same two references — SxrlUBCPpb4 and lMrYN1a0H2M — on different subjects. The Gongbi linework, mineral pigment palette, gilded filigree, and auratic halo carry across both because the style source stayed fixed while the prompt changed. Stable srefs are what make this repeatable instead of a lucky roll.
--sref chooses which style to borrow; --sw controls how hard it asserts itself over your prompt. It runs 0–1000 and defaults to 100. For subject-led shots where you want a light flavoring, 50–100 is plenty. For a strong, brand-locked look, 150–300 is the working band — and the real images in this set push further, sitting at --sw 300 for the cyber portraits and --sw 500 for the Buddhist family, where the style is meant to dominate the rendering.
| Value | Effect in V8.1 |
|---|---|
--sw 0–100 | Subject leads; style is a light wash. Default sits at 100. |
--sw 150–300 | Strong, consistent brand look — the everyday working range. |
--sw 300–500 | Style takes the wheel; the subject bends toward the reference. Where the real images here live. |
--sw 500–1000 | Maximal style pressure; expect the reference to reshape composition, not just finish. |
Reach for --sw when a style is "almost right" but too faint or too loud — lower it when the reference eats your composition, raise it when the look won't come through. The full tuning playbook is in the style weight guide.
Style references travel with a handful of other dials. Order among parameters doesn't matter — they just have to come after your subject text, and the version tag goes last by convention.
| Parameter | What it does in V8.1 |
|---|---|
--stylize | Midjourney's "house aesthetic" strength, 0–1000 (default 100). The real images run it at the maximum, --stylize 1000, leaning into MJ's own polish on top of the sref. |
--chaos | Variety across the four-image grid, 0–100. Low (40) keeps a tight, repeatable family; high (85) spreads the grid for exploration. |
--exp | Experimental aesthetic boost — adds depth and detail richness. The real prompts use modest values like --exp 33; high values can overwhelm a delicate sref. |
--raw | Strips Midjourney's default styling for tighter prompt adherence. Pair it with a strong sref when you want the reference, not MJ's house look, to define the finish. |
--hd | Native 2K output without an upscale step — on by default in V8.1. Remember it does not coexist with --sv 6. |
--sref. Pasting from a doc or chat often converts -- into a "smart" em-dash (—), which Midjourney silently ignores. If a reference seems to do nothing, retype the dashes..png/.jpg/.webp link or an s.mj.run link — not a web page that merely displays the image.--sv. Codes and --sref random only fire on --sv 4 or --sv 6. If a number gets ignored, your style version is the likely culprit.A numeric sref is a coordinate in style space — a random seed that maps one way only: number → style. There is no reverse function. You cannot hand Midjourney an arbitrary image and get back "the sref that made it," and any tool claiming otherwise is bluffing. What's real and useful is the other direction: take an image whose look you want, read its style precisely, then find references that reproduce that feel.
Style-match, don't "extract." Every image above links to its decode — that's what aikizi's decode library does: it reads an image's style DNA, palette, lighting, and finish into the exact tokens you'd need to chase it. You're not reverse-engineering a secret number; you're style-matching against thousands of decoded pieces to find the aesthetic closest to yours, then rolling references toward it. Honest similarity search beats a fake reverse lookup every time.