Midjourney Image-to-Prompt Guide: Turn Any Image Into a Prompt

Midjourney Image-to-Prompt Guide: How to Turn Any Image Into a Prompt That Actually Works

The first time I tried this, I did the dumb thing.

I found a gorgeous portrait on Pinterest, ran it through Midjourney's `/describe` command, copied the first suggestion word-for-word, hit enter, and waited for magic.
What I got back was… fine. Technically competent. Completely different from the image I loved.

That's the moment most people give up on image-to-prompt and decide it "doesn't work." It does work — but not the way you think. Midjourney doesn't reverse-engineer an image. It guesses at it. Your job is to take that guess and turn it into something useful.

This guide is the Midjourney-specific version of the process. If you want the platform-agnostic version — the one that works across Midjourney, DALL·E, Stable Diffusion, and Flux — start with the pillar guide: Image to Prompt: The Ultimate Guide + Free Framework. Then come back here for the Midjourney-only tools, parameters, and quirks.
Let's get into it.

What "Image to Prompt" Actually Means in Midjourney

Here's the thing nobody explains clearly: "image to prompt" in Midjourney is really two completely different jobs, and mixing them up is why people get frustrated.

Job one gives you vocabulary. Job two gives you visual control.
The best results come from using both together. `/describe` teaches you how Midjourney thinks about your image, and reference parameters make sure the model actually looks at it while generating.
Most tutorials only teach you one half. That's why their results look loose.

Method 1: The `/describe` Command (Midjourney Built-In Image-to-Prompt)

This is the native image-to-prompt tool, and it's genuinely useful — as long as you know what you're getting.

How to run `/describe` on Discord

1. Type `/describe` in any Midjourney bot channel or your DM with the bot.
2. Choose the `image` option and upload your file (or use the `link` option and paste a direct image URL).
3. Hit enter.
4. Midjourney returns four prompt suggestions, numbered 1–4.

Under those suggestions you'll see buttons:

• 1️⃣ 2️⃣ 3️⃣ 4️⃣ — generate a grid from that specific prompt
  
• 🔄 — re-roll and get four fresh descriptions
• Imagine all — run all four at once

How to do it on the Midjourney web app

If you're working on midjourney.com instead of Discord, upload your image into the prompt bar (the image icon on the left), then look for the Describe option in the image's menu. Midjourney redesigns this panel fairly often, so if the button has moved, it's usually inside the same dropdown where you assign the image as a style or character reference.
Same output, nicer interface.

What those four results actually are

They're not four attempts at the same answer. They're four different interpretations, usually with different emphasis:

• One leans photographic (camera, lens, lighting terms)
•  One leans artistic (medium, artist-adjacent vibes, art movements) 
• One leans descriptive (subject, colors, composition) 
• One is a bit of a wildcard

Read all four. Don't pick one — harvest all four.

What `/describe` gets right

• Medium and technique. It's good at spotting oil painting vs. 3D render vs. film photography. 
• Color and mood language. It'll hand you words like "muted earth tones," "chiaroscuro," "hazy golden light." 
• Aspect ratio. It usually appends the correct `--ar`. 
• Art vocabulary you didn't know. Honestly the biggest benefit. It'll say "tenebrism" or "isometric" or "cinemagraph" and suddenly you have a search term.

What `/describe` gets wrong (every single time)

• It doesn't know your subject. A photo of your specific product becomes "a sleek modern bottle." A photo of your friend becomes "a young woman." 
• It hallucinates specifics. Camera bodies, lens focal lengths, film stocks, artist names — often invented to sound plausible. 
• It flattens composition. It rarely tells you where things sit in the frame, which is often the single most important thing about the image. 
• It over-indexes on style, under-indexes on structure. You'll get the mood back. You won't get the layout.

So: treat `/describe` output as a word bank, not a finished prompt. That single mindset shift fixes 80% of the frustration.

Method 2: Image Prompts (`--iw`)

Instead of describing the image, you can just… give Midjourney the image. How it works: paste a direct image URL (or upload the image) at the start of your prompt, then add your text after it.

The key parameter is `--iw` (image weight). It controls how much the image influences the result versus your text.

• Lower values (0.25–0.5) → your text dominates, image is a whisper 
• Default (1) → balanced 
• Higher values (2–3) → the image dominates, text becomes a nudge

The exact accepted range shifts between model versions, so if you get an error, drop the number and try again.

When to use image prompts: when you want the overall composition and content echoed — a similar layout, similar objects, similar framing.

When not to use them: when you only want the look. For that, you want style reference instead.

I go deeper into choosing between these in How to Use Reference Images in AI Generation — worth reading if you're regularly working from source material.

Method 3: Style Reference (`--sref` + `--sw`)

This is the one that changed how I work.
`--sref` tells Midjourney: ignore what's in this image, just steal how it looks. Color grading, brush texture, grain, lighting quality, rendering style.

Style weight (`--sw`) controls intensity:

• 0–50 — a faint influence 
• 100 — default, noticeable but balanced 
• 300–500 — strong stylization 
• 700–1000 — the style takes over almost completely

You can stack multiple references:

And you can weight them individually with `::`:

Pro move: `--sref random` gives you a random style seed. When it produces something you love, Midjourney shows you the seed number, and you can reuse it forever with `--sref 1234567890`. It's like building your own private style library without hosting a single image.

When to use `--sref`: brand consistency, series work, matching a visual identity across dozens of images. This is the closest thing Midjourney has to a "house style" button.

Method 4: Omni Reference (`--oref` + `--ow`)

In older versions (V6), character consistency ran on `--cref` and `--cw`. In V7, Midjourney replaced that with Omni Reference — `--oref` — which is broader. It can carry over a person, a creature, an object, or a specific prop into new scenes.

Omni weight (`--ow`) controls how tightly Midjourney clings to the reference:

• (25–75) — loose likeness, lots of stylistic freedom. Good when you're pushing the subject into a very different art style. 
• Default (100) — balanced. 
• High (200–400) — strong likeness, good for photorealistic consistency across a series.

Midjourney's parameter ranges and defaults shift with every model release, so if a number doesn't behave, check the official docs before assuming you broke something.

Rule of thumb I use:

• Want the same person, new scene? → `--oref` 
• Want a different subject, same aesthetic? → `--sref` 
• Want both? → use them together, and expect a few rounds of tuning.

Method 5: Personalization and Moodboards (`--p`)

Once you've rated enough images, Midjourney unlocks personalization — a profile of your taste that you apply with `--p`.

You can also build moodboards: upload a set of images that represent a look, and Midjourney generates a personalization code from that collection. It's essentially `--sref` for a whole aesthetic rather than a single image.

This is the underrated one. If you're producing images for the same brand week after week, spending 30 minutes building a moodboard saves you hundreds of prompt tweaks later.

The 8-Part Midjourney Prompt Formula (That `/describe` Won't Give You)

/describe` gives you adjectives. This gives you structure. Combine both and your hit rate goes way up.

• Subject — who or what, with specific detail
• Action or pose — what they're doing, how they're positioned 
• Setting — where, and what's around them 
• Composition and camera — framing, angle, distance, depth of field 
• Lighting — direction, quality, color temperature 
• Medium and style — photo, oil painting, 3D render, risograph 
• Color palette — dominant colors, saturation, grading 
• Mood — the emotional read

Then parameters at the end.

Watered-down version (what most people write):

Formula version:

Same idea. Wildly different output.

Notice how the second one hits all eight parts without sounding like a robot reading a checklist. That's the goal — write it like you're describing the shot to a photographer over the phone.

My 5-Step Midjourney Image-to-Prompt Workflow

This is the actual process, in order.

Step 1: Describe the image yourself first (60 seconds)

Before touching `/describe`, write down what *you* see. Just bullet points. Subject, layout, lighting, mood.

Why? Because once you read Midjourney's description, you'll anchor to it. Your own read catches things the model misses — especially composition and intent.

Step 2: Run `/describe` and mine it for vocabulary

Now run it. Read all four results. Highlight every phrase that makes you go "oh, that's the word for it."

Ignore everything else. Especially fake camera specs.

Step 3: Merge into one prompt using the 8-part formula

Your bullet points supply the structure. `/describe` supplies the vocabulary. Combine them into one clean prompt.

Keep it under roughly 60 words if you can. Midjourney weights early words more heavily, so front-load your subject and the single most important visual quality.

Step 4: Add parameters and generate a grid

Set your `--ar` to match the reference. Add `--style raw` if you're chasing realism. Add `--sref` or `--oref` if you need the actual image in the loop.

Generate. Look at all four results, not just the one you like.

Step 5: Fix ONE thing per round

This is the discipline that separates people who get there in three rounds from people who flail for thirty.

Look at your grid, identify the single biggest gap between what you got and what you want, and change only the words responsible for that gap.

Wrong lighting? Rewrite the lighting clause. Don't touch anything else.

Use Vary (Subtle) to nudge, Vary (Strong) to explore, and Remix when you need to swap words mid-flight.

Midjourney Parameter Cheat Sheet for Image Matching

The two I reach for most: `--style raw` and `--s`. If your output looks "too Midjourney" — too glossy, too dramatic, too illustrated — those two are almost always the fix.

Worked Example: Matching a Moody Café Portrait

Let's run the whole thing.

Reference image: a woman by a rain-streaked café window, soft grey light, film grain, muted colors, 4:5 vertical.

Step 1 — My own read:

• Vertical, subject on the left third
• Face lit from window on the left, right side falls into shadow. 
• Visible grain, slight halation on highlights 
• Cool grey-blue palette with one warm accent (a coffee cup) 
• Feels lonely, not sad

Step 2 — `/describe` output (paraphrased highlights):

• "cinematic film still, muted color grading" 
• "shallow depth of field, bokeh background" 
• "pensive atmosphere, natural window light" 
• (Also gave me a made-up camera model and a lens spec I ignored)

Step 3 — Merged prompt:

Step 4 — First grid result: lighting was right, grain was right, but the framing was too tight and the mood read as sad rather than contemplative.

Step 5 — One fix: changed "medium close-up" to "medium shot with headroom above the subject" and swapped "pensive" for "calm, self-contained." Left everything else alone.

Round three was the keeper.

Total time: about eight minutes. Compare that to copy-pasting `/describe` and re-rolling forty times.

Troubleshooting: Why Your Midjourney Output Doesn't Match

Copy the Vibe, Not the Artwork

Quick real talk, because this matters.

Midjourney's reference tools are powerful enough now that you *can* get uncomfortably close to someone else's specific piece. Just because the tool lets you doesn't mean it's a good idea — creatively or legally.

The line I hold to: techniques and aesthetics are fair game, specific artworks and living artists' signature output are not.

"Chiaroscuro lighting, muted palette, film grain" is a description of craft. Naming a working illustrator and asking for their exact style is something else.

I've written about where that line sits, how to use references responsibly, and how to build something that's genuinely yours in How to Recreate an Image With AI (Honestly)**. If you're using image-to-prompt for client work, read that one before you ship anything.

Mistakes I See Constantly

Copy-pasting `/describe` output raw. It's a first draft written by something that has never seen your intent. Edit it.

Not matching the aspect ratio. Change the ratio and you change the composition entirely. Set `--ar` before you tune anything else.

Writing 200-word prompts. Midjourney doesn't reward length. It rewards clarity and priority order. Long prompts dilute your key terms.

Changing five things at once. Then you have no idea what worked. One variable per round.

Ignoring the other three grid images. Sometimes image #3 has the lighting you wanted on the composition you didn't. That's information.

Chasing pixel-perfect matches. Midjourney is a re-interpreter, not a photocopier. Aim for "same feeling, same craft level" and you'll be happy. Aim for "identical" and you'll burn a hundred credits.

Skipping the seed. When something is *almost* right, grab the seed and iterate from there instead of re-rolling from scratch.

FAQs

Quick Recap

1. `/describe` gives you vocabulary — not a finished prompt. 
2. Use the 8-part formula to give that vocabulary structure. 
3. Pick the right reference tool: image prompt for composition, `--sref` for style, `--oref` for subject. 
4. Set `--ar` first, `--style raw` and `--s` second. 
5. Fix one thing per round. 
6. Aim for the same feeling, not the same pixels.

Do those six things and Midjourney's image-to-prompt tools stop feeling like a slot machine and start feeling like a workflow.

Midjourney ships updates constantly — parameter names, ranges, and defaults change between model versions. If a flag in this guide throws an error, check the current official documentation, then come back and keep building.