What Metadata Do AI Images Contain?

Local AI tools often write the full prompt, the seed, the model and every setting into the image file, and ComfyUI stores the whole workflow. Some hosted services document C2PA Content Credentials instead. This guide shows, from each tool’s source code, what goes where. Removing metadata does not remove invisible watermarks in the pixels.

Which AI tools write the prompt into the image?

AUTOMATIC1111, Forge, ComfyUI and InvokeAI do it by default, Fooocus does it when its metadata option is on, and so does the reference Flux code from Black Forest Labs. Anyone who gets the original file can read the prompt and the settings.

What each tool writes, read in its source code or documentation
Tool	Format	Where	What
AUTOMATIC1111, Forge	PNG	Text chunk `parameters`	Prompt, negative prompt, steps, sampler, CFG scale, seed, size, model hash and name, LoRA and extension settings
AUTOMATIC1111, Forge	JPEG, WebP, AVIF	EXIF `UserComment`	The same text
ComfyUI	PNG	Text chunks `prompt` and `workflow`, plus any extra key added by nodes	The whole graph as JSON: prompts, seeds, model and LoRA file names, node settings, sometimes local file paths
ComfyUI	AVIF, animated WebP	EXIF tags holding `prompt:` and `workflow:` text	The same JSON
InvokeAI	PNG	Text chunks `invokeai_metadata`, `invokeai_graph`, `invokeai_workflow` (older versions: `sd-metadata`)	Generation settings, graph and workflow
Fooocus	PNG	Text chunks `parameters` and `fooocus_scheme`	Prompt, seed, models, styles
Fooocus	JPEG, WebP	EXIF `UserComment` (parameters), `Software` (“Fooocus v…”), `MakerNote` (scheme)	The same data
Flux reference code	JPEG	EXIF `ImageDescription`, `Software`, `Make`, `Model`	The prompt in ImageDescription; “AI generated;txt2img;flux”, “Black Forest Labs” and the model name in the others

Two details matter for privacy. A ComfyUI workflow is a complete recipe: dropping the PNG back into ComfyUI loads it, with your custom nodes and file names. And Fooocus also writes a log.html file next to the images, with the prompts, which stays on your disk.

The pages on Stable Diffusion metadata and on the ComfyUI workflow explain how to read, remove or stop writing these fields.

Where is metadata stored in an image file?

In blocks set aside from the pixels, which every image format allows. Four kinds carry what AI tools write: PNG text chunks, EXIF, XMP and C2PA manifests.

PNG text chunks (tEXt, zTXt, iTXt) are pairs of a key and a text, compressed or not. Any program can add one with any key, which is why each AI tool picked its own names.
EXIF comes from cameras: maker, model, date, exposure, GPS position. AI tools reuse some of its fields, such as UserComment, ImageDescription, Software or MakerNote, to store their settings in JPEG and WebP files.
XMP is an XML packet used by Adobe software and many others. It can hold the creator tool and the IPTC digital source type, a standard way to say how an image was made: trainedAlgorithmicMedia means “created using generative AI” and compositeWithTrainedAlgorithmicMedia “edited using generative AI”.
C2PA manifests, also called Content Credentials, are signed records of how the image was made, stored as JUMBF data (C2PA 2.2, annex A). The C2PA page explains what they contain and when to keep them.

Photos add their own layer: camera, lens, date and often the GPS position. See photo metadata and GPS.

What do hosted AI generators write?

The vendors that document it describe C2PA Content Credentials saying that the image was made by AI, by which service and when, rather than the prompt. Two examples are published:

Microsoft: images from Azure OpenAI carry a manifest signed by Microsoft, with the description “AI Generated Image”, a software agent such as “Azure OpenAI DALL-E” or “Azure OpenAI ImageGen”, and the time of creation.
Black Forest Labs: JPEG and PNG images from the FLUX API carry Content Credentials with an AI-generated flag, the claim generator, the UTC time, and the model name and version, and no prompt, negative prompt, API key, account ID or IP address.

For ChatGPT, Midjourney, Gemini and Firefly, we have not inspected a real file yet. Published reports differ, and some come from sellers of removal tools, so this guide states nothing about them. We publish a page for a generator only after inspecting a real file from it.

Does removing metadata remove AI watermarks?

No. An invisible watermark is written into the pixels themselves, not into a metadata block, so no metadata tool removes it, PromptWipe included. Visible logos are part of the picture too.

SynthID: Google says SynthID embeds imperceptible signals in AI-generated content, and that the Gemini app can check an uploaded image for it.
DWT-DCT, from the open-source invisible-watermark package: Stability AI’s SDXL reference scripts call it, the Flux reference code calls it on every image it saves, and the diffusers SDXL pipeline adds it by default when the package is installed.

So an image can carry no metadata at all and still be recognised as AI-generated. PromptWipe is a privacy tool for what the file says about you, not a way to hide how an image was made.

How can you check what your own image contains?

Open it in the PromptWipe viewer: it lists every block of the file and shows the prompts, settings, EXIF fields and Content Credentials it finds, without sending the image anywhere.

Keep in mind that a copied, pasted or screenshotted image has often lost its metadata already, and that many social platforms re-encode uploads, which often drops it too. To know what a file you are about to share carries, check that file itself, then share the cleaned copy.

Sources: AUTOMATIC1111 images.py; ComfyUI nodes.py and nodes_images.py; InvokeAI, media metadata; Fooocus private_logger.py; Flux util.py; IPTC digital source type; C2PA 2.2 specification; Microsoft Learn; Black Forest Labs help; Google, image verification in Gemini; Stability AI generative-models; diffusers.