The best style reference is the one that holds a look across a whole series — not a single lucky frame. Every "top sref codes" roundup you have seen lists numbers like --sref 1093847 and promises a vibe. Those numeric codes are real, but they are random coordinates in Midjourney's style space; they drift between model versions and tell you nothing about why an image looks the way it does. This page does the opposite. Below are seven images generated on Midjourney V8.1, each shown with its real prompt, its real image-based style references, and the style DNA we decoded out of the pixels. Two of them belong to families — one reference set, several finished images, one coherent aesthetic. That is the part worth learning.
Which V8.1 are we on? Midjourney V8.1 landed on April 30, 2026 and became the default model on June 10. It is the fastest version to date and ships sharpness gains that are most visible exactly where style references and moodboards live. Everything below is real V8.1 output — prompts end in --v 8.1.
Midjourney's --sref accepts two very different things, and conflating them is the root of most confusion. A numeric style code (--sref 2841) points at a fixed region of the model's internal style library — portable, shareable, but abstract. An image reference (--sref https://s.mj.run/<id>) points at an actual picture and tells the model "borrow the look of this." Every example on this page uses image references, because that is how a real working aesthetic gets built: you find an image whose feel you want, lock it in, and re-run your concepts through it.
One honest caveat up front. Those s.mj.run links are Midjourney-hosted and personal — they resolve to a specific creator's uploaded reference and can expire. We reveal them so you can read complete, real V8.1 prompts, not because the URL is a durable asset to bookmark. The durable thing is the workflow and the decoded style DNA. More on that at the end.
Here is the whole argument in one move. Three different finished images — a Himalayan thangka, a Nihonga moon scene, and a Tang-dynasty Gongbi figure — all came out of the same two image references: SxrlUBCPpb4 and lMrYN1a0H2M. The subjects change, the model version and the references do not, and the result reads as one painter's hand across the set. That consistency is the entire reason image references exist.
Swap the concept to a moonlit Nihonga scene and the references carry the ultramarine wash, the monumental scale, and the calligraphic edge straight across — same hand, new subject. The --exp 33 on both adds a touch of experimental flair without breaking the lineage.
On --sw in V8.1. Style weight sets how hard the model grips your reference. The range is 0–1000 and the default is 100; the --sw 500 above is a deliberate, heavy pull toward the source — earned here because the references and the concept already agree. Drop it toward 100 or lower when you want the prompt to lead and the reference to merely tint. Tune it per source, not by habit. We unpack this fully in the --sw deep dive.
The same trick, a different aesthetic. These two — a speculative afrofuturist portrait and a tech-noir engraving — share the references I8ZEy5X1Xco and xVNj3CzP9Lo. Notice the --raw and the lower --sw 300: with a high-chaos cyber look you want the references to set the palette and texture while leaving the model room to invent the hardware.
That right-hand image keeps the duotone and the etched-line discipline even as the subject shifts to a hard profile portrait. Same references, same chromatic polarity, different framing — a series, not a one-off.
Not everything needs two references. A single strong source can carry a whole image, especially when the concept is doing heavy lifting. Here are two standalone V8.1 results that each lean on one reference and one decisive parameter choice.
Read those two against each other and the parameter logic becomes obvious. The chimera wanted --weird and a lighter grip; the deity wanted a heavy --sw and zero weirdness. The reference is the constant — the parameters are how you steer it.
| Parameter | What it does in V8.1 | Used above |
|---|---|---|
--sref <url> | Borrow the look of one or more reference images | Every example |
--sw | Style weight: how hard the model grips the reference (0–1000, default 100) | 300–500 |
--chaos | Variation across the 4-up grid | 40–85 |
--exp | Experimental flair; still available in V8.1 | 33, 41 |
--weird | Pushes toward unusual, off-distribution form | 33 |
--raw | Reduces default styling, raises prompt adherence | cyber · chimera · deity |
What about --oref? Omni Reference and its --ow weight are not supported in V8.1 — attaching an omni image drops the job onto the older omni model instead. For locking a character across images, that path still exists; for locking a style, --sref is the V8.1 tool. We separate the three reference types in sref vs cref vs iref.
You do not memorize these URLs. You build your own. The workflow that produced every family above is short and repeatable:
1. Roll references. Generate your concept with no --sref at all, or with a single throwaway reference and high --chaos. You are fishing for one frame whose feel is right, not whose subject is right. Sharpness and SREF behavior are noticeably better in V8.1, so this fishing step lands faster than it used to.
2. Promote the winner. When a frame feels right, use it as the reference for the next batch — drop its image URL into --sref and re-run your real concepts. This is exactly how the Buddhist family was built: one reference set, then subject after subject.
3. Save winners as a set. Keep your two or three best references together. A pair like SxrlUBCPpb4 + lMrYN1a0H2M is more stable than either alone — two references average toward a consistent house look instead of one image's quirks.
4. Tune --sw last. Once the references are fixed, the only dial left is how hard the model grips them. Lower it for invention, raise it for fidelity. Nothing else. The how-to guide walks the full sequence with screenshots.
Be clear about the boundary, because the internet is full of tools claiming otherwise. There is no way to point at any picture on the web and extract "the sref that made it." A numeric style code is a random coordinate, and an image reference is a private URL — neither is recoverable from finished pixels. Anyone promising to reverse a sref from an upload is selling you a guess.
What is real is reading the style DNA. aikizi's decode looks at an image and names what is actually there — the mineral pigments, the chiaroscuro, the crosshatched engraving, the chromatic polarity — the same tokens shown on the pills above. Click any image on this page and you land on its full decode. From there you style-match: rebuild that look with your own references and parameters, the honest way. Browse hundreds of decoded V8.1 examples in the gallery, or read style codes explained for the full mental model.
Decode a style →This is part of a four-part series on Midjourney style references in V8.1:
- Midjourney style codes explained — the pillar: what sref really is
- How to use sref codes in Midjourney V8 — step by step
- sref vs cref vs iref — style, character, and image references compared
- Midjourney --sw style weight — tuning how hard a reference grips
- The decode gallery — hundreds of real V8.1 images, decoded