Midjourney Sref Codes Are Version-Bound: A Library Strategy That Survives

If a style reference you saved months ago now renders something completely different, nothing is broken – your library is version-bound and the ground moved. Midjourney’s numeric sref codes are tied to the style-algorithm version (--sv) they were captured under, the algorithm was rewritten in mid-2025, and the V8 family’s default (--sv 7) simply ignores bare numeric codes from earlier eras. I maintain a catalog of thousands of captured srefs, and reconciling it against the V8 family produced the compatibility matrix and the library strategy below.

TL;DR

  • A numeric sref code is not a style. It is a pointer into a specific style algorithm. Change the algorithm version and the same integer renders a different aesthetic.1
  • The break line is the mid-2025 algorithm rewrite: codes captured before it render faithfully only under --sv 4 on V7. Codes from the later era need --sv 6 in the V8 family – under V8’s default --sv 7, bare numeric codes are ignored.
  • Image-URL srefs are the durable form. A reference image is interpreted fresh by whatever algorithm runs it; a numeric code is frozen against a dead one. New library entries should be images first, codes second.
  • --sref random – the discovery engine – runs under --sv 4 or --sv 6 only, and pairs with Draft Mode on V8.1, not V8.2.2
  • The current --sv 7 system is about 4x faster and 4x cheaper than its predecessor, so regenerating your working set beats living in compatibility mode.3
  • Full parameter coverage lives in the Midjourney guide; the V7-to-V8 migration map covers the rest of the surface.

What an Sref Actually Transfers

The --sref parameter transfers aesthetic qualities from a reference – color palette, lighting style, artistic technique, mood, and compositional tendencies – without transferring subject matter. Strength runs through --sw (0-1000, default 100): under 50 is a whisper, 50-150 is a balanced transfer, 150-300 asserts the style, and above that the style dominates the prompt. Multiple references blend: --sref [url1] [url2] averages aesthetics into combinations no single image holds.

Srefs reward distinctive sources. A reference with a strong visual identity – consistent palette, recognizable technique – transfers cleanly; a generic photo transfers mush. This matters for library-building: the codes worth cataloging are the opinionated ones.

The Version-Binding Problem

Here is the part that costs people their libraries. When you capture a numeric sref code, the integer is meaningful only to the style algorithm that minted it. Midjourney has revised that algorithm across --sv 4, --sv 6, and now --sv 7 – and the mid-2025 rewrite was deep enough that codes from before it render a completely different style unless explicitly pinned to their home version. This is the single most common reason an old sref “looks wrong” in the V8 family: the code did not drift, the default interpreter changed underneath it.

The failure is silent. No warning, no error – just a render that is aesthetically unrelated to what you cataloged, under the same code you have used for a year.

The Compatibility Matrix

Four rules cover every case:

Your sref Where it works faithfully
Numeric code, recent era V8 family with explicit --sv 6 (the default --sv 7 ignores bare numeric codes)
Numeric code, pre-rewrite Only under --sref <code> --sv 4 --v 7 – it does not survive into V8
Image URL V8 family under --sv 6 or --sv 7 – interpreted fresh, no era binding
--sref random --sv 4 or --sv 6 only; pair with Draft Mode on V8.1 for cheap hunting2

The asymmetry in that table is the strategy. Numeric codes are brittle across versions; image references are not, because an image is re-interpreted by whatever algorithm receives it. The cleanest V8-native path is image-based srefs.

Building a Library That Survives

The practice that emerged from maintaining a large sref catalog through two algorithm transitions:

1. Capture the image, not just the code. When a random hunt or a generation produces an aesthetic worth keeping, save the source image alongside the numeric code. The code is a fast handle for today’s algorithm; the image is the style’s permanent record. When the next --sv arrives, your codes will wobble and your images will not.

2. Tag every code with its --sv era at capture time. A bare integer in a notes file is a future mystery. 1234567890 (sv6, captured 2026-03) is usable forever, because the matrix above tells you exactly how to run it.

3. Hunt on the cheap tier, render on the good one. --sref random under Draft Mode on V8.1 is the discovery engine – 24-image grids at half cost, sampling aesthetics you would never prompt for deliberately.2 Catalog the hits, then apply them to real work as image references on V8.2.

4. Deduplicate and cohort. A library past a few hundred entries develops near-duplicates – the same aesthetic family captured five times under different codes. Periodically cluster by visual similarity and keep the strongest exemplar of each cohort; a library’s value is its range, not its count.

5. Regenerate the working set, archive the rest. The current --sv 7 system runs about 4x faster and cheaper than its predecessor.3 For the twenty styles you actually use, re-establish them natively (from the saved source images); keep the long tail in compatibility mode with era tags rather than migrating everything.

Srefs vs Moodboards: The Division of Labor

Approach Best for
--sref One-off style from a specific reference – per-project, per-shot control
--p moodboard A consistent personal aesthetic applied across everything

They compose: a moodboard supplies your baseline taste while an sref steers an individual render, and both ride the same --sv 7 infrastructure in the V8 family.3 The library strategy differs accordingly – moodboards are tended like a garden (curated additions over time), srefs are cataloged like a spice rack (labeled, dated, reached for by name).

FAQ

Why does my old Midjourney sref code look different now?

Because numeric codes are bound to the style-algorithm version that created them, and Midjourney rewrote the algorithm in mid-2025. Codes from before the rewrite render faithfully only under --sv 4 on V7; later codes need explicit --sv 6 in the V8 family. The change is silent – same code, different interpreter, different style.

Do numeric sref codes work in Midjourney V8?

Only with an explicit style version: --sref <code> --sv 6. Under V8’s default --sv 7, bare numeric codes are ignored. Pre-2025 codes are effectively V7-only (--sv 4 --v 7). Image-URL references work natively across --sv 6 and --sv 7, which is why a durable library stores images, not just integers.

What is –sv in Midjourney?

The style-version parameter: it selects which style-reference algorithm interprets your sref. --sv 4 and --sv 6 are earlier systems kept for backward compatibility; --sv 7 is the V8-era baseline, roughly 4x faster and cheaper than its predecessor.3 A numeric code only means what its home algorithm says it means.

Should I use an sref or a moodboard?

Sref for steering a specific render toward a specific reference; moodboard (--p) for applying your accumulated personal taste everywhere by default. They combine in one prompt for layered control – baseline from the moodboard, per-shot direction from the sref.



  1. Midjourney, Version documentation – the V8-family compatibility surface for style references, and the merged V8.1/V8.2 column this post’s matrix is checked against. Full working detail in the Midjourney guide

  2. Midjourney, Draft Mode for V8.1 – Draft Mode’s 24-image grids at half the fast-hours of a standard V8.1 job; the --sref random sampler pairs with it on V8.1, and the rebuilt compatibility chart does not carry Draft Mode to V8.2. 

  3. Midjourney V8 Alpha SREF update (March 2026) – the --sv 7 baseline: approximately 4x faster and 4x cheaper than the previous style system, supporting --hd, --p, --stylize, and --exp, with moodboards defaulting to the same system. 

Articles connexes

Midjourney V8 vs V7: The Prompt Anatomy That Actually Changed

V8 executes prompts literally where V7 interpreted them, and half of V7's control surface quietly changed underneath. Wh…

8 min de lecture

Midjourney V8 Prompting: What Changed from V7

Midjourney V8 prompting: personalization-first workflows and what V8.1 Alpha changed. Default HD, image prompts back, st…

21 min de lecture

Sign in with Apple Sends Four Notifications, Not Three

Apple's announcement names three server-to-server notification types. The API defines four. Here is the full contract, a…

13 min de lecture