What does /describe actually do?

/describe asks Midjourney to look at an uploaded image and write four text prompts that could plausibly have produced it. It is a reverse-engineering tool — it guesses descriptive language, not the exact original prompt.

Which tokens are worth keeping?

The valuable parts are reusable style descriptors — medium (oil painting, 3D render), lighting (golden hour, rim light), mood, and named aesthetics. Subject-specific nouns and generic filler like "highly detailed" add little and can be dropped.

Can /describe recover the exact original prompt?

No. It reconstructs likely descriptive language, not the literal prompt, seed, or parameters. Two images made the same way can yield different /describe outputs, so treat it as inspiration, not a recovery tool.

How do I reuse extracted style tokens?

Combine the style, lighting and medium tokens with your own new subject. For example, take "cinematic lighting, 35mm film, muted palette" from one image and apply it to a completely different scene to carry the look across images.

Is my pasted text uploaded?

No. The tool parses and categorises your /describe text entirely in the browser. Nothing is sent anywhere.

Midjourney /describe Reverse Engineer Guide

Midjourney /describe reverse-engineer guide

Midjourney’s /describe looks at a reference image and writes four candidate prompts that could have produced it. The real value isn’t the whole prompt — it’s the reusable style tokens hidden inside. Paste a /describe result and this tool splits it into phrases, buckets them into style, lighting, medium and composition, and flags generic filler to drop.

How to mine /describe output

/describe reconstructs language, not the literal original prompt — so don’t copy it wholesale. Instead, harvest the transferable parts:

Medium — “oil painting”, “3D render”, “35mm film”, “vector art”.
Lighting — “golden hour”, “rim light”, “soft studio lighting”.
Mood / aesthetic — named styles, eras, and atmospheres.
Composition — “close-up”, “wide shot”, “centered”, “rule of thirds”.

Discard the subject nouns (they tie the prompt to that specific image) and the generic filler (“highly detailed”, “trending”, “8k”) that adds noise without steering the look.

Tips for building a reusable style

Keep a token library of medium + lighting + mood combos you like.
Swap the subject, keep the style to carry a look across different scenes.
Combine tokens from multiple /describe runs to invent a hybrid aesthetic.
Test with a --sref style reference too — it often beats text tokens for pure style transfer.