Skip to content
Creativo Labs

AI tool

Image to Prompt

Reverse-engineer a prompt from any reference image. Choose whether to capture the whole image, just its style, or just its subject — then get it written to your target model's conventions.

Reference image

Drag and drop, browse, or paste with Ctrl/Cmd + V. This is the image the prompt gets written from.

Prompt options

How much of the reference to capture, and which model the wording is written for.

Comma-separated descriptors with parameter flags

Describe subject, composition, lighting and style — everything needed to reproduce it.

Added as an --ar flag.

0 / 300

Change something about the reference, e.g. “same lighting but a winter scene”.

Recent prompts

Saved in this browser only. Never uploaded anywhere.

About reverse-engineering a prompt

Sometimes you can see the look you want but cannot name it. This tool reads a reference image and writes a prompt that describes it, in the conventions of whichever model you intend to run.

Three focus modes decide what gets described. The whole image captures everything; style only takes the treatment — lighting, palette, medium, mood — and leaves the subject out so you can substitute your own; subject only does the reverse.

When this is the right tool

  • You have a reference whose treatment you want to reuse with a different subject.
  • You are building a set that needs one consistent visual language across many images.
  • You cannot articulate why an image looks the way it does and want vocabulary for it.
  • You want to understand the lighting or medium in a picture, not just copy it.

How to use it

  1. Upload a reference image

    It is downscaled in your browser before being sent for analysis, which is faster and makes no difference to the description.

  2. Choose the focus

    Whole image, style only, or subject only. Style only is the mode to reach for when you are building a consistent set.

  3. Choose the target model

    The description is written to that model's conventions, so it can be pasted straight in rather than reworded first.

  4. Generate, then edit

    Treat the result as vocabulary rather than a finished prompt — swap in your own subject, and keep the clauses that describe the look.

Requirements and limits

Input
JPG, PNG and WebP up to 10MB. A web-sized copy is plenty — extra resolution does not make the description better.
Never names artists or brands
The description works in terms of technique, lighting and medium rather than naming a living artist or a trademark. That is both a licensing matter and a more reusable description.
What is sent
A downscaled copy of the image, proxied to Google and discarded after the response. It is not stored or logged.

Common questions

What does it actually give me?

A generation prompt describing the reference image, written to the conventions of the model you plan to run it in — Midjourney, Flux, Ideogram, Firefly or ChatGPT Images. It is a description of the image, not the original prompt that made it, which no tool can recover.

What is the difference between the three focus options?

Recreate describes everything — subject, composition, lighting, style. Style only describes the treatment and leaves the subject out, which is what you want when you are borrowing a look for your own subject. Subject only does the reverse.

Will the result be identical to the reference image?

No, and it should not be. The prompt captures what is describable — subject, framing, light, palette, medium. Re-running it produces something in the same territory, not a copy.

Can I use it on someone else's image?

Use it on images you have the right to work from. The tool deliberately never names living artists or brands, and describes technique and genre instead, so the output does not lean on someone's name — but that does not make a copyrighted reference free to imitate.

Is my image uploaded to a server?

It is sent to Google's Gemini API through this site's own route handler, because reading the image is what the model does. It is not stored by this site, and there is no gallery or account holding it afterwards.

What file types and sizes work?

JPG, PNG and WebP up to 10MB. A web-sized version is enough — extra resolution does not make the description better.

Guides on this

Longer answers to the questions this tool raises.

Related tools