Image to prompt: read the hidden data and colours in a picture

What this tool does and does not do

What hidden data a picture can hold

A picture file is more than its pixels. Different programs add different notes, and each lives in its own place in the file. This is what the tool looks for.

Anyone can write or change most of these notes, so they show what a file says, not proof of how it was made. The tool shows only what it finds and says "not found" for the rest.

How to get the prompt from a Stable Diffusion PNG

1. Use the original PNG. Take the file straight from the app's output folder, not a screenshot and not a copy saved as JPEG.
2. Drop it into the box above, or paste it with Ctrl+V.
3. Open "Prompt data from an AI app". The prompt, the negative prompt and the settings (steps, sampler, CFG scale, seed, size, model) are listed with Copy buttons.
4. To use the words in Gemini or ChatGPT, tap "Use as my subject". It keeps the words and drops weights such as (word:1.3) and extra-network tags such as <lora:name:0.8>, which only work in the app that wrote them.

A ComfyUI PNG holds the full workflow too. Drop that PNG into ComfyUI itself to reload the graph.

Colour palette to prompt

The tool makes a small copy of your picture in the browser and finds five main colours in it. It uses a median cut followed by a few k-means steps, with no randomness, so the same picture always gives the same palette. Each colour gets the nearest of about fifty plain names, such as terracotta, sage or navy. Colour names read the same in every app, so the prompts use the names, and the hex codes stay on the swatches and the card.

Brightness, contrast, saturation and warmth become words like low-key, soft contrast, muted and warm. They are estimates from the pixels, not a measurement of the scene, and they describe how a picture looks, not what is in it. You can switch both the colour names and the exposure words off.

Why the prompt is often missing

Instagram, WhatsApp and X usually strip metadata from pictures you post or send, and they have not published exact rules, so it can vary. That is why screenshots and downloaded posts usually have no prompt data, no camera data and no location. Turning a PNG into a JPEG, or opening and re-saving it in an editor, can also drop the data. If you want the prompt of a picture, ask for the original file or look at where it was first shared.

The same fact works in your favour: sharing an original photo as a file can share its location. Make a clean copy first.

Remove location and camera data from a photo

The clean copy is drawn again on a canvas in your browser and saved as a new file. Camera data, GPS, text chunks, XMP and Content Credentials are left behind. This recompresses the image: JPEG and WebP files lose a little quality, PNG stays lossless, and the colour profile is dropped, so some wide-gamut colours can shift slightly. A photo over about 16 megapixels is scaled down so that phones can save it. After saving, the tool reads the new file again and tells you if anything is left.

How the formats were checked

The file layouts here were checked against each app's own source code or documentation on 15 October 2026. Apps change, so if a layout stops working, the source is the place to look.

• Automatic1111 web UI: how the "parameters" text is written and read (modules/images.py, modules/infotext_utils.py): the parameters text (a PNG text chunk; the EXIF user comment for JPEG and WebP), and the NovelAI layout it also reads.
• ComfyUI docs: workflow metadata in saved files: the prompt and workflow keywords, and the EXIF form used by animated WebP.
• InvokeAI: PNG text chunks and the core metadata fields: the invokeai_metadata, invokeai_graph and invokeai_workflow chunks and the metadata field names.
• Fooocus: metadata written to PNG (modules/private_logger.py, modules/meta_parser.py): the parameters and fooocus_scheme chunks, in JSON or Automatic1111 style text.
• C2PA reference SDK: where a manifest sits in JPEG, PNG and WebP files: where a Content Credentials block sits: a JPEG APP11 segment, a PNG caBX chunk, a WebP C2PA chunk.
• IPTC: the digital source type value for AI-generated media: the digital source type value trainedAlgorithmicMedia in XMP.
• Midjourney docs: Aspect Ratio and Raw: --ar written as width:height at the end of the prompt, with whole numbers.
• OpenAI image prompting guide: the scene, subject, details, constraints order, the word photorealistic, and camera terms being read loosely.

The Gemini, ChatGPT and Midjourney prompts follow each app's own guidance. For Midjourney the only settings used are --ar and --raw, the two confirmed in its Raw and Aspect Ratio pages. To check a Content Credentials signature, the Content Authenticity Initiative runs a verify tool; it uploads your file, so read its privacy terms first.

More for creators

Prompt tools and guides for the same kind of pictures.

AI photo prompt generator Cinematic prompt guide: lens, light and film Make AI photos look real Why AI photos look fake Stable Diffusion prompts Colour grade prompts