QWE AI Academy You’ve found an image with the exact aesthetic you need. Maybe it’s a vintage poster, a moody photograph, or an illustration style your client wants. Now the question: how do you teach Midjourney to recreate it?
Two methods. /describe extracts text prompts from images. –sref clones the visual style directly. Most tutorials treat these as separate features. They’re not – they solve different parts of the same problem, and the real workflow uses both.
Why /describe Alone Isn’t Enough
The /describe command is Midjourney’s reverse-engineering tool. According to Midjourney’s official docs, upload an image and it returns four text prompt suggestions that attempt to describe what it sees.
Sounds perfect for recreating styles, right?
Not quite. /describe is designed to be “inspirational and suggestive” – it won’t precisely replicate your reference image. Run the same image through /describe multiple times and you’ll get different suggestions each round. That’s by design.
Here’s the bigger issue: the prompts often name specific artists. “In the style of Keith Negley” or “reminiscent of Martin Ansin.” These artist references heavily influence the output – but if you’re creating commercial work, using copyrighted artist names introduces licensing gray areas. Remove the artist names? The output drifts away from your reference image.
This is the trap nobody mentions in basic tutorials.
When –sref Wins (and When It Doesn’t)
Style Reference (–sref) takes a different approach. Instead of converting your image into words, it feeds the image directly into Midjourney’s style system.
What it captures: colors, textures, lighting mood, medium (watercolor, oil, digital, etc). What it ignores: specific objects, composition, subject matter. Per the official Style Reference guide, think of it as copying the “vibe” without copying the content.
The advantage: no artist names, no licensing concerns, and tighter style consistency across multiple images. You control how strong the influence is with –sw (style weight), which ranges from 0 to 1000. Default is 100.
Pro tip: Start with –sw 200 for your first test. Default (100) is often too subtle, and 500+ can overpower your text prompt. Find your sweet spot between 150-300 for most use cases.
But –sref has limits. It won’t help you if the reference image’s subject is what you need to recreate – only its aesthetic treatment. If you’re trying to match a specific pose, object, or scene layout, –sref won’t get you there. That’s when Image Prompts (not covered here) become necessary.
The Hybrid Workflow: Using Both Together
Here’s the technique most tutorials don’t cover: use /describe to generate vocabulary, then feed the original image as –sref while using that vocabulary in your prompt.
Step-by-step:
- Upload your reference image and run /describe
- Review the four prompt suggestions – look for descriptive terms (not artist names) that capture what you want
- Pull out 5-8 useful descriptors: lighting terms, material descriptions, mood words
- Craft a new prompt using those terms to describe your actual subject
- Add your reference image as –sref at the end
Example: you want to create a product photo in the style of a moody editorial shot. /describe suggests “dramatic chiaroscuro lighting, muted earth tones, film grain texture, shallow depth of field.”
Your hybrid prompt:
sleek wireless headphones on marble surface, dramatic chiaroscuro lighting, muted earth tones, shallow depth of field --sref [your reference image URL] --sw 180The –sref handles color palette, lighting mood, and texture. The text prompt (informed by /describe) handles subject and composition. Together they’re stronger than either alone.
Actually, there’s no UI button for this workflow – you’re manually bridging two separate features. But it works because Midjourney’s documentation confirms Style References can combine with text prompts. Most users just don’t realize /describe output makes better text prompts for this combo than writing from scratch.
Practical Setup: Web vs Discord
The interface matters because /describe behaves differently depending where you use it.
On the Web:
Right-click any image in your Create or Organize pages, select “Describe Image.” You’ll see modular blocks – Subject, Descriptor – not complete sentences. These are building blocks you drag into your prompt bar. It’s more flexible but requires more assembly work.
To add –sref on web: click the image icon in the Imagine bar, upload your reference, drag it into the Style Reference section (not the Image Prompt area – different functions).
On Discord:
Type /describe, upload your image or paste its URL. You’ll get four complete text prompts you can click to run immediately. The output is more verbose and includes those artist name references we talked about.
For –sref in Discord: upload your reference image first (it generates a Discord CDN URL), then type your prompt followed by –sref [paste that URL].
Which platform? Web UI is cleaner for iterating on style references. Discord is faster if you just need /describe suggestions to learn prompting vocabulary.
The V7 Version Confusion Nobody Explains
Here’s a gotcha that’ll waste your time if you’re using community-shared sref codes.
Midjourney updated the Style Reference system on June 16, 2025. The new algorithm (–sv 6, now default) is smarter about interpreting style and avoids “subject leakage” where elements from your reference image bleed into unrelated generations.
Great. Except all those sref code libraries you find online? Most were created before June 2025. If you try –sref 1234567 and it looks nothing like the preview image, it’s because that code was generated under the old system.
The fix: add –sv 4 to your prompt to use the old V7 model, or switch to V6 entirely. But here’s what’s annoying – most sref code websites don’t mention this. They’ll show you a beautiful example, you’ll copy the code, and your result will be wildly different until you figure out the version mismatch.
Check the date on any sref resource you use. Post-June 2025? You’re fine with defaults. Earlier? Append –sv 4 to the code or expect surprises.
When to Skip These Methods Entirely
Not every style problem needs /describe or –sref. Sometimes you’re fighting the wrong battle.
Skip /describe if your reference image is super niche – traditional crafts, regional art styles, technical diagrams. Midjourney’s training data might not have vocabulary for it, so /describe will return generic terms that don’t help.
Skip –sref if you need exact replication. The official docs are clear: Midjourney uses references “as inspiration to guide new creations, not to copy them exactly.” If your client needs pixel-perfect matching, you need the Editor tool or a different approach entirely (inpainting, img2img, or just Photoshop).
Both methods struggle with text-heavy reference images – logos, typography, infographics. Midjourney can’t reliably reproduce readable text regardless of which method you use.
What Actually Controls Style Strength
You’ve seen –sw mentioned (style weight for –sref). Values from 0 to 1000, default 100. But how do the numbers actually feel?
- 0-50: Barely perceptible. Use this if you want just a color palette hint.
- 100-200: Noticeable but balanced. Your text prompt still dominates.
- 300-500: Strong style influence. Reference image’s aesthetic takes over.
- 600-1000: Maximum style adherence. Your text prompt fights for relevance.
One thing: –sw behaves differently with sref codes (those numeric IDs) versus sref image URLs. With V7, style codes respond more dramatically to weight changes than images do. So –sw 500 with a code might feel like –sw 700 with an image URL.
Test in increments. Generate the same prompt at –sw 100, 250, and 500. See where your sweet spot is before committing to variations.
The Describe + Sref Handoff
Let’s map out the full workflow with a concrete scenario.
You found a product photo with that perfect “minimalist Scandinavian product photography” look. Clean, soft shadows, muted palette. You need to create similar images for different products.
Step 1: Run /describe on the reference. It returns terms like “soft diffused lighting, neutral backdrop, pastel color grading, shallow focus, negative space composition.”
Step 2: Extract the useful descriptors. “Soft diffused lighting” and “pastel color grading” are keepers. “Shallow focus” too. Ignore any artist names.
Step 3: Write your subject-specific prompt: “ceramic coffee mug on white surface, soft diffused lighting, pastel color grading, shallow focus, negative space composition”
Step 4: Add the reference image as –sref. Upload it to Midjourney (web or Discord), grab the URL, append: –sref [URL] –sw 220
Step 5: Generate. The sref handles the color mood and lighting quality. Your text (informed by /describe) handles composition and subject.
Why this works: /describe taught you the language Midjourney understands for that style. –sref ensures the color and lighting feel matches even if your subject is completely different.
Next action: grab a reference image you’ve been struggling to recreate. Run this hybrid workflow once. You’ll immediately see why text-only or sref-only approaches leave gaps.
FAQ
Can I extract the sref code from an image I didn’t create?
No. Sref codes are generated randomly by Midjourney when you use –sref random, but there’s no reverse lookup to extract a code from an existing image. If you want to reuse a style from someone else’s image, you need the actual image file to feed into –sref as a URL, not a numeric code. The only way to get reusable codes is to generate images with –sref random, note the code Midjourney assigns (shown in the job parameters), and save it for later.
Why does my –sref output look nothing like the preview in an sref code library?
Version mismatch. Midjourney updated the Style Reference algorithm on June 16, 2025. Codes generated before that date use the old system (–sv 4); codes after use the new default (–sv 6). If you’re trying a code from an older library, add –sv 4 to your prompt to match the original system. Alternatively, check if the library specifies which version the code was created under – most don’t, which is exactly the problem.
How many sref images can I use in one prompt, and do they blend or compete?
You can use multiple image URLs with –sref (just separate them with spaces) or multiple numeric sref codes. They blend – Midjourney averages the style characteristics rather than picking one. You can even weight individual references by adding ::number after each URL (higher numbers = more influence). Example: –sref URL1::2 URL2::1 gives URL1 twice the style weight of URL2. Practical limit is 3-4 references before the output becomes muddled; two is the sweet spot for intentional style fusion.