Midjourney: style with --sref (Style Reference)
When the brief says «like this illustration» or «same palette as this poster», stacking adjectives often returns a generic catalog look. Midjourney Style Reference (`--sref`) applies an image’s aesthetic — palette, grain, brushwork, contrast — without forcing you to copy its objects or people. That is different from an Image Prompt (composition and content) and from Character / Omni Reference (same person or object across shots). Official docs: https://docs.midjourney.com/hc/en-us/articles/32180011136653-Style-Reference and Creating on Web: https://docs.midjourney.com/hc/en-us/articles/33390732264589-Creating-on-Web. In about 18 minutes you build a repeatable flow: pick a ref, attach it, write a clear shot, lock a grid before vary or upscale (midjourney-variaciones-upscale).
The problem it solves
People drag random Pinterest boards into Style Reference and complain about subject leakage, or mix --sref with a product photo as an image prompt and get an unreadable hybrid. Another trap: describing style in ten adjectives when one clean reference would do — or using --sref for something you could state in one line («flat vector, two colors») and paying GPU for little control. Without separating look from shot, every grid is a random experiment.
What you will learn
- Tell Style Reference (--sref) from Image Prompts and character refs (Character / Omni)
- Pick a clear style image (not a chaotic collage)
- Attach Style Reference on the web and with --sref on Discord
- Combine --sref with a shot prompt (--ar, subject, light)
- Tune --sw when needed and know when to skip --sref
- Lock a grid composition before variations or upscale
Before you start
- Paid Midjourney account with access to https://www.midjourney.com or Discord
- A job with a defined subject — ideally after midjourney-prompts-que-funcionan
- A style image you may use (yours, licensed or public domain); no third-party brands without permission
- 18 minutes and a pad to copy the final prompt with parameters
Steps 1
Split three jobs: look (--sref), content (image prompt), identity (Character / Omni)
Style Reference answers «what does it look like?»: color, texture, light quality, medium (watercolor, soft 3D, hard flash). It does not promise the same character or composition. Image Prompts (URL at the start on Discord, Image Prompt slot on web — https://docs.midjourney.com/hc/en-us/articles/32040250122381-Image-Prompts) push composition and content toward the reference photo. Character Reference (`--cref` on Discord in classic flows) and Omni Reference (`--oref`, Omni slot on web — https://docs.midjourney.com/hc/en-us/articles/36285124473997-Omni-Reference) put the same person or object into new scenes; on V8 many flows use Edit Model on the web, but the rule stays: identity ≠ style. If you only want palette and grain, use Style Reference — do not drop the model photo into Image Prompt «just in case».
What you should see: A note with three columns: «look → Style Reference / --sref», «shot/content → text + sometimes Image Prompt», «same face/object → Omni/Character/Edit Model». Nothing generated yet.
Steps 2
Pick a style reference you can read in two seconds
Good ref: one illustration, poster or still with a homogeneous look — limited palette, same texture across the frame, not three styles in one file. Bad ref: four-photo collage, screenshot with UI, or a photo where «style» is really the subject (a specific person). Crop to show aesthetic only (background, brushwork, grain). Valid formats: .png, .webp, .jpg per docs. If the ref is too literal (very recognizable hero product), Midjourney may still pull object cues even though --sref targets look; V7+ reduces subject leakage (https://updates.midjourney.com/style-references-for-v7/), but a clean ref is still cheaper than ten failed grids.
What you should see: One file (or one numeric Style Creator `--sref` code) noted as «style ref». No five URLs on the first test.
Steps 3
Attach Style Reference on the web (Create)
Open Create at https://www.midjourney.com. Click the image icon beside the Imagine bar, upload or pick from Uploads (10 MB max per Creating on Web). Drag the image into the Style Reference slot — not Image Prompt or Edit Model. Type a minimal shot prompt: «empty ceramic vase on shelf, eye-level, soft daylight from left, plain wall --ar 4:5». Generate a 2×2 grid. The ref does not replace subject or camera: you should recognize your shot wearing the reference’s visual skin. For an internal code instead of an image, Style Creator on the web builds numeric `--sref` codes (https://docs.midjourney.com/hc/en-us/articles/41308374558221-Style-Creator); same flow: code or image for look, text for the shot.
What you should see: Ref thumbnail in the Style Reference slot under the Imagine bar. A 2×2 grid with your job and a look close to the ref.
Steps 4
Same flow on Discord with --sref URL
On Discord, `/imagine` with a publicly reachable image URL (upload to Discord and copy link if needed). Shape: [shot prompt] --sref https://… --ar 4:5. Multiple refs: separate URLs with spaces; per-URL weights with `URL1::2 URL2::1` (Style Reference docs). Do not paste the URL only at the start — that is an Image Prompt, not Style Reference. Compare one web grid and one Discord grid with the same prompt and ref so your team can repeat the flow on either surface.
What you should see: Bot message with visible prompt, --sref and --ar in the queue. Grid consistent with the web test.
Steps 5
Combine --sref with a locked shot (subject, light, --ar)
Rewrite the prompt like midjourney-prompts-que-funcionan: concrete subject, angle, light, exclusions, then parameters. Example with a watercolor ref: «single pear on white plate, slight cast shadow, eye-level, one window soft light, no hands, no text --ar 4:5 --stylize 80 --sref [your URL or web slot]». Low --stylize keeps obedience to the shot; --sref supplies the aesthetic. Do not repeat «watercolor dreamy artistic» in text if the ref is already watercolor — you compete with yourself. If style fights the shot (hard-flash ref + soft fog prompt), fix light in text or change ref; do not chain ten generations.
What you should see: Four cells sharing the ref look but respecting subject and framing. At least one usable cell without rewriting everything.
Steps 6
Tune --sw only when the ref is too weak or too strong
Style weight `--sw` runs 0–1000; default 100 (Parameter List and Style Reference on docs.midjourney.com). Raise --sw (e.g. 200–350) when the shot «eats» the look. Lower --sw (e.g. 40–80) when the ref crushes the subject or oversaturates. On Discord: `--sref URL --sw 250`. On the web, append `--sw` to the prompt if no slider is visible. Do not touch --sv (sref algorithm version) unless you reuse old codes and docs ask for `--sv 4`. If --sw 100 is balanced, stop tuning and lock composition.
What you should see: Two comparable grids: lower --sw (more shot obedience) vs higher --sw (closer to ref). Your note on which you ship.
Steps 7
When NOT to use --sref (and when text is enough)
Skip Style Reference if: (1) you can name the look precisely and do not need an external visual («isometric line icon, two flat colors, no gradient»); (2) the ref is a random photo where you cannot separate look from subject; (3) you blend two opposite style refs «to see what happens»; (4) the client banned external references. Use text + low --stylize instead. If the job is recurring character, do not crank --sw on --sref — switch to Omni/Character/Edit Model for your version. If the job is copying an ad layout, a legal image prompt + written shot is more honest than a style ref scraped from someone else’s campaign.
What you should see: Checked list: «can I state the look in one line?» «homogeneous ref?» «identity vs look?». Written decision: sref yes/no.
Steps 8
Lock composition and hand off to variations / upscale
Pick the winning cell and write one sentence why it wins («cell 3: ref palette match, pear centered, readable shadow»). Copy the full prompt with --sref (or web ref) and --sw if used. Stop changing style unless the brief changes. Next: midjourney-variaciones-upscale — Vary Subtle for detail, Upscale Subtle once when composition is locked. For motion, runway-primer-clip with the HD JPG; --sref already did its job on the still. Do not re-upload the same ref «just in case» in Runway — animate the approved frame.
What you should see: Saved prompt, noted ref, marked cell on the grid without upscale yet (or one upscale if you already locked). Session ready to vary.
Real use cases
Campaign with a fixed brand palette
One internal still (not the logo) in Style Reference + product in text. --sw 120 if the brand is very saturated. Lock the grid before upscale.
Editorial illustration
Homogeneous brush ref, prompt with a concrete news subject. No Image Prompt of someone else’s art. Vary Subtle only on the winning cell.
Team on Discord
Director posts approved ref URL and a prompt template with --ar. Same --sref in thread as on the designer’s web Create.
Common mistakes
Using a collage moodboard as the only --sref
Four styles fight. Midjourney averages noise. One ref, one look.
Confusing Style Reference with Image Prompt
URL at the start on Discord copies composition, not just look. --sref goes at the end with a written shot.
Asking for style in adjectives and --sref at once
Redundancy pushes the default «pretty». Let the ref speak or drop adjectives.
Raising --sw to fix a vague prompt
If the subject never appears, it is not missing --sw: it is a weak shot. Return to midjourney-prompts-que-funcionan.
Upscale before the look is locked
You spend GPU on a look that still wobbles. Lock a cell on the SD grid, then vary/upscale.
Close
Style Reference (`--sref`) transfers an image or code’s aesthetic; it does not replace a written shot or clone characters (that is Omni/Character/Edit Model). The flow that works: homogeneous ref → Style Reference slot or --sref URL → subject, light and --ar in text → fine --sw if needed → lock a cell → vary or upscale once. When you can describe the look in one clear line, sometimes low stylize text alone beats a ref.
Next steps
- Tighten the base shot in midjourney-prompts-que-funcionan if the subject still does not read.
- Run Vary and Upscale in midjourney-variaciones-upscale on the cell you locked here.
- Animate the HD still in runway-primer-clip when composition is fixed.
Free guide
Want the 15-tools guide?
We send it free. We will also ping you when the next guide in this path goes out.
Unsubscribe whenever you want. We do not sell the list.