Midjourney --sref Explained: Style References, Codes and Weights (2026)

TL;DR: --sref transfers the look of a reference image or a numeric style code onto a new Midjourney generation. Colours, medium, texture, lighting. Not subjects, and not characters. Strength is set with --sw (0 to 1000, default 100), the algorithm with --sv. Works on version 6 and later. Verified August 27, 2026.

What is --sref in Midjourney?

--sref is the style reference parameter. You hand Midjourney a picture or a code, and it borrows that picture's aesthetic for whatever you actually asked for.

Midjourney's own Style Reference article puts it like this: a style reference "doesn't copy objects or people, just the overall style—like colors, medium, textures, or lighting—helping you achieve a consistent visual theme." That sentence is the whole feature. Everything else is syntax. On midjourney.com, drag your upload into the Style Reference slot in the images panel. In Discord, put --sref at the end of the prompt followed by a public image URL.

a lone fisherman mending nets on a stone pier at dawn --sref https://example.com/my-reference.jpg --v 8.2

The published constraints are short. Your file has to end in .png, .gif, .webp, .jpg or .jpeg, and it has to already be online, so a desktop file needs hosting first. You also always need text alongside the reference: "for your prompts to work, you need a text prompt in combination with your Style Reference."

What does --sref copy, and what does it refuse to copy?

It copies treatment. It does not copy content, and it does not copy people.

This is where most --sref frustration comes from, and Midjourney addresses it with paired good and bad examples. The bad ones: the look of this image but a dog, copy this style and make a bunny. The good ones: detailed portrait of a dog, ballpoint pen sketch of a bunny. The rule underneath is describe the picture you want, not the edit you want. The reference is already doing its job silently, and the prompt's only job is the subject.

Its third best practice is counterintuitive enough to quote verbatim: "Keep text prompts simple - Avoid adding style words that might conflict with your reference image's look." If your reference is a soft gouache illustration and your prompt says "hyperrealistic 8k octane render", you have given the model two contradictory briefs and it will split the difference badly. Our Midjourney style modifiers reference covers that vocabulary; when a style reference is doing the work, most of it should come out.

How do you control style strength with --sw?

--sw sets how hard the reference pushes. Midjourney publishes a range of 0 to 1000 with a default of --sw 100.

brutalist concrete library interior, wide shot --sref 1284523901 --sw 400 --v 8.2

brutalist concrete library interior, wide shot --sref 1284523901 --sw 40 --v 8.2

Run those two and the difference is obvious: at 400 the reference is close to dominating palette and composition, at 40 it is a tint. There is no documented "correct" value, and anyone quoting one is quoting their own taste.

One published nuance saves an hour of confusion. Midjourney notes that "If using V7, you may notice that style weight has more impact when used with style codes than with images." So --sw is blunter on uploaded references than on codes, at least on V7. The docs do not extend that to V8.1 or V8.2 either way.

What are Midjourney sref codes, and where do you get them?

A style reference code is a number that points at a style in Midjourney's internal style library. You use it exactly where you would use an image URL.

a ceramic teapot on a windowsill, morning light --sref 2954650647 --v 8.2

Midjourney documents three routes to a code, and rules out a fourth:

1. --sref random. Submit it, and the word resolves into a real code you can copy.
2. The Style Explorer. On the Explore page, click the Styles tab. Sort by random or popularity, fuzzy-search terms like "photographic" or "anime", preview a code's gallery, hit Try Style, or Like it to save it. Liking a style code does not affect your Personalization profiles.
3. The Style Creator. A midjourney.com-only tool that builds a custom code from styles you pick and reject across successive rounds. Midjourney's guidance: most styles stabilise after 5 to 10 rounds, rounds 10 to 15 add detail, past round 15 changes are subtle. Previews cost GPU time, and the tool renders in V7.

What you cannot do is turn your own upload into a code. Midjourney is explicit: "You cannot create a style code based on an uploaded image." If you want your photograph's look, you pass the photograph every time.

How does --sref random actually work?

--sref random applies a random code from the library, then rewrites itself. Midjourney: "After submitting your prompt, 'random' converts to a style code." What happens next depends on what else is in the prompt, and the split is documented precisely:

• Rerun, reroll, or variations keep the same style code. Good for iterating on a look you just found.
• Permutations, --repeat, or V8.1 draft mode give each image a different style code. Good for exploring.

a single origami crane on black --sref random --v 8.2

a single origami crane on {black, bone white, oxidised copper} --sref random --v 8.2

The second prompt is a permutation, so every one of those three renders gets its own style. That is the cheapest style-discovery loop Midjourney publishes.

The feature's history explains its shape. Midjourney introduced --sref random on April 22, 2024 as "a test of a underlying algorithm". When the V7 style system shipped on June 16, 2025, the announcement said "There's a new --sref random! the new styles should be much broader than before." On June 25, 2026 it reached V8.1 draft mode, where a single draft prompt returns 24 images in 24 different styles.

One limit is easy to miss: Midjourney notes that "Sref random and sref codes are only compatible with --sv 4 and --sv 6." Pin --sv 1, --sv 2, --sv 3 or --sv 5 and codes stop working.

What is --sv, and which version are you actually on?

--sv selects which style reference algorithm interprets your reference. It is the least-known of the three dials, and the likeliest explanation for a style code that used to look right.

On V6, Midjourney describes all four: --sv 1 is "the original style reference algorithm, more 'vibey'", --sv 2 the second iteration, --sv 3 an updated and more vibey --sv 1, --sv 4 an updated --sv 2. That set arrived on March 21, 2024 for both Midjourney V6 and Niji V6. On V7 there are six, of which the docs name two: --sv 6 is the default, and --sv 4 is "the old V7 sref model (prior to June 16, 2025)". Versions 1, 2, 3 and 5 exist in the count but carry no published description.

That June 2025 date is why old codes go strange. Midjourney's own note: "older style codes may not produce the same styles anymore. To use old codes, add --sv 4 to your prompt (uses the old V7 model) or switch to V6."

art nouveau botanical poster, muted ochre and sage --sref 3403639244 --sv 4 --v 7

Can you use more than one style reference at once?

Yes, and Midjourney documents no cap. On the web you can select multiple images in the images panel; in Discord you separate URLs with spaces. Codes mix the same way, and Midjourney says you "can also mix style codes by using more than one, and combine images with style codes."

a rain-slick Tokyo alley at night --sref 2077548100 3403639244 --sw 250 --v 8.2

Weighting individual references is where the documentation gets uncomfortable, and it is worth being honest about rather than papering over.

The Style Reference article says: "When using Midjourney in Discord, individual Style References can also be assigned different weights: --sref URL1::2 URL2::1 URL3::1", and links to Multi-Prompts & Weights for the syntax. But that article says: "Multi-prompts work with Midjourney versions 1, 2, 3, 4, Niji 4, 5, Niji 5, 6, Niji 6, and 6.1." And the Version chart marks Multi-Prompting as unsupported on V7, V8.1 and V8.2.

Both pages were revised in late July 2026, both are live, and Midjourney reconciles neither. The plain reading is that per-reference :: weighting is a V6-era Discord feature the Style Reference page still advertises without a version scope. We are not going to guess which page is right. Test it before you build a V8.2 workflow on it.

How does --sref interact with the other style controls?

Midjourney ships four overlapping style systems, and confusing their dials is the second most common --sref failure.

Two things follow from that table that people get wrong constantly.

--stylize is not the style reference dial. Midjourney's Stylize article describes it as a general artistic-freedom slider, and the Personalization and Moodboards articles both say specifically that stylize governs how much of your personalization or moodboard is applied. Neither says it governs --sref. The documented dial for a style reference is --sw.

Competition for influence is documented for --oref, not for --sref. The Omni Reference article warns that "If you're using high stylize or --exp values you may want to also use a higher Omni Reference Weight, as these will all compete for influence." There is no equivalent sentence for style references. Community advice about --stylize fighting --sref may be right in practice, but it is consensus, not documentation, and this page will not launder it into a rule.

Raw is the one clean combination. Midjourney describes --raw as switching off its default aesthetic auto-pilot, on every version from 5.1 onwards, which makes it the closest thing to "just use my reference and nothing else". We covered its limits in does --style raw still work in Midjourney.

technical cutaway of a diesel locomotive, clean linework --sref 1284523901 --sw 300 --raw --v 8.2

Does --sref work with --cref, --oref and Niji?

Yes to all three, with version caveats that matter more than the syntax.

Midjourney states that "Style Reference can be combined with Character References/Omni References and Image Prompts", and the Omni Reference article adds that you can use --oref "in combination with Personalization, Moodboards, stylize, and Style References." Style plus subject in one prompt is supported, not a hack.

The catch is that the subject parameters are version-locked in opposite directions. The Version chart marks Character Reference as V6 only, Omni Reference as V7 only, and both as unavailable on V8.1 and V8.2, where the Omni Reference rows read "(Uses V7)". Style Reference is the only one of the three ticked across all four. If you need a consistent character and a consistent style, you are not on the current default model. Post character consistency in Midjourney with --cref and --oref covers that trade in depth.

a woman in a red raincoat crossing a bridge --sref 2954650647 --sw 200 --oref https://example.com/character.jpg --ow 250 --v 7

Niji is the other half of the answer. Midjourney's Niji V7 announcement of January 9, 2026 lists "improved anime coherence, prompt understanding, text rendering, and sref performance" as the release headlines. Spellbrush's own Niji 7 post goes further: under a heading titled "Give sref a Try!", it says "We've added backwards-compatibility for most of the features in this release. (Yes, even the weirdly, specifically requested --sv 4 flag)".

a girl with short green hair reading on a train, anime screenshot --sref 3403639244 --niji 7

Sixteen --sref prompts worth keeping

Swap the codes for your own. The numeric examples come from Spellbrush's published --sref random walkthrough; treat them as syntax illustrations, not recommendations.

Brand and series consistency

product hero shot of a matte black water bottle on seamless backdrop --sref https://example.com/brandboard.jpg --sw 300 --ar 4:5 --v 8.2

product hero shot of a folded grey merino scarf on seamless backdrop --sref https://example.com/brandboard.jpg --sw 300 --ar 4:5 --v 8.2

blog header illustration, abstract data flowing between two towers --sref 2954650647 --sw 150 --ar 16:9 --v 8.2

Locking a look before you scale it

isometric illustration of a coffee roastery --sref random --v 8.2

isometric illustration of a {coffee roastery, bicycle workshop, record store} --sref random --v 8.2

isometric illustration of a bicycle workshop --sref 2077548100 --sw 100 --v 8.2

Dialling strength deliberately

oil portrait of an elderly bookbinder at his bench --sref 1284523901 --sw 800 --v 8.2

oil portrait of an elderly bookbinder at his bench --sref 1284523901 --sw 60 --v 8.2

Mixing two styles

a lighthouse in fog, long exposure --sref 2954650647 2077548100 --sw 250 --v 8.2

Reference plus subject

a man with silver hair and round glasses in a greenhouse --sref 3403639244 --sw 180 --oref https://example.com/face.jpg --ow 200 --v 7

Reaching back to an older style model

art deco travel poster, alpine funicular --sref 3403639244 --sv 4 --v 7

art deco travel poster, alpine funicular --sref 3403639244 --v 6

Style reference plus Raw

architectural photograph of a concrete chapel, overcast --sref https://example.com/reference.jpg --sw 350 --raw --v 8.2

Anime and Niji

a fox spirit in a bamboo forest at dusk --sref 3403639244 --niji 7

Wide formats, where the ceiling changes

panoramic matte painting of a canyon city --sref 2077548100 --sw 200 --ar 12:1 --v 8.2

editorial spread background, soft gradient mesh --sref random --sw 120 --ar 3:1 --v 8.2

Why is my --sref not working?

Six documented causes, in the order they actually occur:

1. No text prompt. A style reference on its own is not a prompt. Midjourney requires both.
2. Style words fighting the reference. Remove the adjectives and let the reference lead.
3. Wrong file type or an unhosted file. It must be a live URL ending in .png, .gif, .webp, .jpg or .jpeg.
4. An old code on a new model. Codes made before June 16, 2025 may not reproduce. Add --sv 4 or drop to --v 6.
5. A moodboard in the prompt. Moodboards disable both --sw and --sv.
6. An --sv that does not support codes. Random and codes work only on --sv 4 and --sv 6.

If none of those apply, the seventh cause is that your reference has no strong style to transfer. A flat product photograph on white gives --sref almost nothing to carry over. That is the feature working on a weak input, not a bug, and putting the look into words first usually fixes it faster than raising --sw. Our guide to turning an image into a reusable prompt is the shortest route there.

How to verify any of this in sixty seconds

Every parameter claim on this page ages. docs.midjourney.com returns 403 to a plain browser fetch, which is expected, not a dead link. Its Zendesk JSON API answers normally:

# Style Reference: --sref, --sw, --sv, sref random, the multi-image weighting note
curl -s "https://docs.midjourney.com/api/v2/help_center/en-us/articles/32180011136653.json"

# Version: the feature compatibility chart, release dates, current default model
curl -s "https://docs.midjourney.com/api/v2/help_center/en-us/articles/32199405667853.json"

# Moodboards: the --sw and --sv exclusions
curl -s "https://docs.midjourney.com/api/v2/help_center/en-us/articles/39193335040013.json"

Each response carries an updated_at field. Check it before you trust the body, and check the body before you trust this page. For dated announcements, updates.midjourney.com is directly fetchable and publishes sitemap-posts.xml.

Worth saying plainly, because plenty of pages blur it: Prompt Architects does not generate images. We generate the prompt. Midjourney renders it. What we are good at is the part that reliably breaks, which is carrying the same subject description, style vocabulary and parameter block across fifty renders instead of retyping it from memory each time. Save the --sref block once, make the subject a variable, and the code, the weight and the version stop being things you have to remember.

Sources and dates

Every claim above was checked against the vendor's own pages on August 27, 2026, via the Zendesk JSON API for docs.midjourney.com and HTTP/1.1 for nijijourney.com. Both hosts return 403 to an ordinary fetch.

Midjourney help centre, all revised July 2026: Style Reference (July 27, the primary source for --sref, --sw, --sv, codes and random) · Version (July 24, the compatibility chart) · Stylize · Moodboards · Personalization · Omni Reference · Character Reference · Style Creator · Multi-Prompts & Weights · Seeds · Raw.

Dated announcements: Random Style References (April 22, 2024) · New Style Refs (March 21, 2024) · Style references for V7 (June 16, 2025) · V8 Alpha (March 17, 2026) · Random styles in draft mode (June 25, 2026) · Version 8.2 (July 24, 2026) · Changelog 8/20/26 (August 21, 2026) · Niji V7 (January 9, 2026).

Spellbrush, on nijijourney.com: Welcome to Niji V7 (stamped January 8, 2026) for the --sv 4 note and the statement that --cref is not supported on Niji 7, and How to use sref random (April 26, 2024) for the random-resolves-to-a-number behaviour and the example codes reused above.

Stated rather than guessed. Midjourney publishes no --sv values or default for V8.1 and V8.2, no maximum number of style reference images, no format specification for style codes, no statement that --stylize competes with --sref, and no reconciliation between the Style Reference page's :: weighting example and the Version chart's "no multi-prompting on V7 or V8". Where this page says "not documented", that is the finding, not a gap in the research.