Use Midjourney style + character refs

Midjourney has the strongest style-consistency toolkit of any consumer image generator. Most users never touch it because the parameter names are obscure (--sref, --oref, --sw, --ow, --sv) and the docs are scattered. This skill consolidates the workflow so your agent can deliver a coherent image set on the first try instead of generating 40 one-offs.

There are three tools, and they do different things. Mixing them up is the #1 mistake.

Tool	Parameter	What it carries	What it does NOT carry
Style Reference	--sref URL or --sref CODE	Colors, medium, texture, lighting, "vibe"	Subjects, characters, composition
Omni Reference (V7+)	--oref URL	A specific person, object, vehicle, or creature	Style, lighting, composition
Image Prompt (legacy)	URL at start of prompt + --iw N	A blend of content and style	Fine-grained control

Use --sref when the user wants a look. Use --oref when the user wants a thing in the picture. Use both together when they want both. Image prompts are blunt, reach for them only when neither of the above fits.

Style Reference (--sref)

Syntax in Discord: prompt text --sref <URL_or_CODE> On midjourney.com: drag the image into the Style Reference slot in the Imagine bar.

Three input forms:

1. Image URL. --sref https://example.com/inspiration.jpg. The image must already be hosted online; if the user has a local file, tell them to host it on Discord first.
2. Style code. --sref 1234567890. Numeric codes from MJ's internal style library. Browse them on midjourney.com/explore under the Styles tab.
3. Random. --sref random. MJ picks a style code and substitutes it in. Pair with --repeat to fan out across many styles.

Combine multiple refs with spaces: --sref URL1 URL2 URL3. Combine with weights to bias toward one: --sref URL1::2 URL2::1.

Style Weight (--sw)

--sw N controls how strongly the reference influences output. Range 0–1000, default 100.

• --sw 50. light influence, prompt text leads
• --sw 100. balanced (default)
• --sw 400. heavy influence, reference dominates
• --sw 1000. extreme, prompt text barely matters

The default is usually wrong. Try --sw 200 first when you want a clear style transfer.

--sw is not compatible with Moodboards (see Personalization profiles instead).

Style Reference Version (--sv)

The --sref algorithm has multiple versions. For V7:

• --sv 6. default, current
• --sv 4. old V7 model from before June 2025 (use if reproducing old looks)

For V6: --sv 1 through --sv 4 exist with different aesthetics. Don't tell the user to use these unless they ask, defaults are fine.

Prompting around --sref. the trap

The biggest mistake is telling MJ what to do with the reference. Don't do this:

❌ the look of this image but as a dog --sref URL ❌ copy this style and make a bunny --sref URL

The prompt text should describe the content you want, not the relationship to the reference:

✅ detailed portrait of a golden retriever --sref URL ✅ ballpoint pen sketch of a bunny --sref URL

If the user gives you a reference and asks "make a dog in this style," translate it before generating: drop the meta-language, keep the subject.

Omni Reference (--oref)

Syntax: prompt text --oref <URL> Only one image, you can't stack --oref URLs.

Puts a specific person, character, or object into the generated image. The reference image is what you want to appear; the prompt text describes the scene around it.

illustration of a young woman with short gray hair drawn by a comic book artist, sitting in a 1980s diner --oref https://example.com/character.jpg --ow 100

Omni Reference Weight (--ow)

--ow N controls how faithfully the reference is reproduced. Range 1–1000, default 100.

• --ow 50–100. loose interpretation; MJ will alter details to fit the scene
• --ow 200–400. moderate fidelity; use this for character consistency across a series
• --ow 500+. high fidelity but unpredictable unless paired with high --stylize or --exp

Keep --ow under 400 unless the user specifically asks for near-photographic accuracy.

Omni Reference gotchas, the agent MUST know these

1. 2× GPU cost. Omni Reference doubles the per-image GPU time vs regular V7. Factor this into iteration budget.
2. Not compatible with Fast Mode, Draft Mode, or Conversational Mode. If the user is on Draft, --oref won't work, they need to switch off --draft first.
3. Not compatible with inpainting, Vary Region, Pan, or Zoom Out. To edit an --oref result, the user must open it in the midjourney.com Editor and remove the --oref + --ow params.
4. V7 only. Doesn't work on V6 or earlier.
5. Style reinforcement is critical. If you want a style different from the reference image, mention the desired style both at the start and end of the prompt. repetition matters because --oref pulls hard toward the reference's style.

Combining --sref and --oref

Yes, this works, and it's the magic move. The combo lets you:

• Put a specific character (--oref) into a specific aesthetic (--sref)
• Build a whole shoot's worth of consistent character-in-style imagery

woman in a green coat walking through fog at night, cinematic, moody --sref STYLE_CODE --oref CHARACTER_URL --sw 200 --ow 200 --ar 16:9 --v 7

Tip: lower both weights (--sw 150 --ow 200) when combining, because they compete for influence.

Workflows the agent should know

Workflow 1: Consistent character across 10+ images

1. User uploads or describes a character. Generate one strong character portrait first using --oref with their reference at --ow 200.
2. Lock the resulting image as the --oref for all subsequent shots.
3. Vary the scene, not the character: change prompt text + --ar, keep --oref URL --ow 200.
4. If style starts drifting, add --sref of the first generated image to lock both.

Workflow 2: Consistent style across a series with different subjects

1. User has a mood image or finds one in the Style Explorer.
2. Lock it as --sref with --sw 200.
3. Generate each scene with different prompt text but the same --sref + --sw.

Workflow 3: Style discovery for a new project

1. Run the same prompt 10 times with --sref random.
2. Each output gets a different style code. MJ writes it back into the prompt.
3. The user picks favorites; copy the resulting --sref CODE for future use.
4. Pair with --repeat 4 if the user has Standard+ plan to fan out in fewer clicks.

Workflow 4: Editorial pivot, same character, different mood

1. Use --oref to lock the character.
2. Swap --sref codes between shots to change mood (warm/cool, soft/hard, photo/illustration).
3. Keep --ow 200 constant so character stability doesn't change with mood.

GPU-cost reminder

Setup	Cost per image job
Base V7 image	1 GPU-min
--sref only	1 GPU-min (no surcharge)
--oref (any weight)	2 GPU-min (2× regular)
--sref + --oref	2 GPU-min

If the user is on Basic ($10/mo, ~200 image jobs/month), 50 Omni Reference shots = 100 GPU-min = half the month's budget. Push them toward Standard ($30/mo) if they're doing character work seriously.

When NOT to use these tools

• Single-image generation, no consistency need. use the midjourney-v7-prompts skill, save the GPU.
• Text rendering / typography. --sref doesn't preserve text, and --oref won't put a logo in the right place. Suggest Photoshop compositing.
• Highly literal product photography. --sref makes everything moody and editorial. Use plain V7 with --raw --stylize 50.

What to deliver to the user

For consistency-driven requests, return:

1. The base prompt template (with placeholders for the user's --sref / --oref URLs)
2. The shot list. one prompt per scene, all sharing the locked refs
3. Estimated total GPU cost based on number of shots × per-shot rate

Example:

**Locked refs (use across all shots):**
- Character: --oref <USER_CHARACTER_URL> --ow 200
- Style: --sref <USER_STYLE_URL> --sw 200

**Shot 1, diner interior**
> woman in a green coat sitting at a diner counter, neon outside, rain on windows --oref ... --sref ... --ow 200 --sw 200 --ar 16:9 --v 7

**Shot 2, exterior street**
> woman in a green coat walking past closed shops, wet pavement, sodium streetlight --oref ... --sref ... --ow 200 --sw 200 --ar 16:9 --v 7

**Shot 3, close-up portrait**
...

Estimated cost: 6 shots × 2 GPU-min (Omni Reference) = 12 GPU-min ≈ $0.40 on Standard plan.