AI Image Metadata: Prompts, Settings and Provenance
Learn what AI image metadata can reveal about prompts, seeds and Content Credentials, how to inspect an export, and how AI watermarks differ.
AI-generated images can carry ordinary metadata, generation settings, or signed information about how they were made. A PNG may contain the prompt and seed used to create it. Another export may contain Content Credentials that record an AI action. What you can learn depends on the application, its export settings, and what happened to the file after it left the generator.
What metadata can an AI-generated image contain?
Record or signal | What it can tell you | Where to look |
|---|---|---|
Ordinary metadata | EXIF can record orientation, software, dates or GPS. XMP and IPTC can carry captions, credits and other descriptive fields. PNG text can carry comments and application data. These fields are generally editable. | Inside the file, in structures that vary by format. |
Generation parameters | A prompt, seed, model or checkpoint name, sampler, or workflow. Each exporter decides what to write. | Often in PNG text when a local application saves it. Job details may also live in an account, project or API response. |
C2PA Content Credentials | Signed claims about creation, edits, ingredients or the use of AI, depending on what the signer recorded. | In a manifest associated with the image. The manifest can be embedded or stored externally. |
Invisible watermark | A signal such as SynthID that a compatible detector can look for. | In the image data. |
Platform AI label | A notice attached to a post. For example, Meta says its AI info label can draw on industry-shared signals or an uploader's disclosure. | In the platform interface. Meta's labeling policy |
An EXIF software tag may describe the application that last saved the file. If an editor wrote that field, it tells you about that step in the file's history. The same applies to a caption or software name in XMP.
Where are AI prompts and seeds stored?
A prompt is the instruction supplied to an image generator. A seed helps initialize the generation process. Depending on the application, those values may appear in exported metadata, in a generation job, or both. A prompt visible in a gallery and a seed returned through an API describe the job. Inspect the downloaded file to see what traveled with it.
PNG stores textual information in tEXt, zTXt, and iTXt chunks. With metadata enabled, AUTOMATIC1111 v1.10.1 can save generation settings under parameters in a PNG. ComfyUI's built-in Save Image node can write an execution graph under prompt and additional supplied records such as workflow. The prompt entry can be a graph with many settings, including a seed, rather than a single instruction. W3C PNG specification · AUTOMATIC1111 PNG writer · ComfyUI Save Image documentation
A published ComfyUI img2img example contains both prompt and workflow in PNG text. Its execution data includes a seed, sampler and step count. AiMetaOff displays supported text records without running their workflows. Pinned ComfyUI example
How exporters differ
ChatGPT Image. OpenAI documents Content Credentials and SynthID for supported image output. In a published Content Authenticity Initiative sample attributed to ChatGPT, the companion manifest report lists an AI creation action in the image's recorded history. Inspection of that PNG found an embedded credential and no prompt in the supported file records reviewed. The exact download path is unknown, and local verification of this sample returned an overall invalid result. OpenAI provenance documentation · CAI sample manifest
Gemini. Google documents SynthID for supported images it generates or edits. Checking a downloaded file for embedded metadata and checking for a watermark are separate steps. Gemini export guide · Gemini Apps Help
Adobe Firefly. Adobe says it automatically applies Content Credentials to assets where all pixels were generated using Firefly, such as a Firefly Text to Image asset. Adobe also documents that credentials may be stored in its cloud. Firefly export guide · Adobe Content Credentials overview
Midjourney. The website documents controls for copying a prompt, job ID or seed. Those controls tell you what is available in the job interface. Check a specific download for embedded records. Midjourney export guide · Midjourney web documentation
Stable Diffusion. The application saving the image determines which generation records are written. Compare the Stable Diffusion PNG examples for AUTOMATIC1111 and ComfyUI.
FLUX. A published FLUX.1 dev ComfyUI example contains workflow and execution text. In BFL's hosted FLUX 3 Image workflow, the documented result can include an expanded prompt in JSON alongside the image URL. BFL also documents signed credentials for final images in its own hosted workflows and says the credential excludes the text prompt. FLUX export guide · BFL API workflow · BFL credentials documentation
What do Content Credentials establish?
A C2PA manifest contains a signed claim and associated assertions. For example, an assertion may record that a tool generated the image or that an editor performed an action. A content binding uses cryptographic hashes to check that the manifest matches the asset data covered by the binding. A signature check checks the signature on the claim. Signer trust depends on the certificate chain and the verifier's trust policy. Even with successful checks and a trusted signer, the assertions are still statements made by that signer, and the recorded history can be incomplete. C2PA technical specification
AiMetaOff reports container detection and available offline verification results separately. A report can show the claim signature, assertion hashes and content binding when the embedded manifest can be read and the corresponding checks are available. AiMetaOff reports signer trust as unverified and leaves remote manifests, online trust lists and revocation lookups outside its inspection. A credential can be present even when a verification check fails or is unavailable.
How do invisible watermarks and platform labels work?
Google DeepMind's SynthID puts an imperceptible signal into the image data. It is designed to survive some transformations, including forms of lossy compression. A compatible detector is needed, and detectability depends on the image and the detector's coverage. AiMetaOff's metadata inspection and cleanup leave watermark analysis outside their scope. Google DeepMind's SynthID overview
A screenshot creates a separate file and usually leaves the source file's metadata behind. Editing, converting, or sharing through a platform can preserve, replace or discard metadata, and editing may invalidate an existing content binding or result in a newly signed credential. A platform can still show an AI label based on information it has about the post. OpenAI's provenance overview · Meta's labeling policy
Can metadata identify the AI generator?
A model name in editable text gives you a lead. For stronger evidence, examine a credential's AI action, whether it refers to creating or editing the image, the signer, and the validation results. The signer, the app that saved the file and the model used to generate it can all be different. If the image lacks supported metadata or a readable credential, its origin remains unknown from that inspection. A human-edited image can contain AI-generated elements, and a fully generated image can have no readable provenance.
How to inspect AI image metadata
Open AiMetaOff's Inspect File tool and choose the original download if you have it. It reads the file locally in your browser.
Review Text and XMP for readable generation records or descriptions. Check EXIF for supported fields such as software or GPS, and review IPTC findings where supported. JPEG APP13 findings identify the resource block. Individual IPTC tags are outside the viewer's decoding scope.
Read the Content Credentials findings separately. Check what the file actually claims and which checks were available. An unsupported or incomplete check leaves that part of the file unresolved.
AiMetaOff supports common JPEG, static PNG and static WebP variants, with narrower support for single-image AVIF and limited HEIC. Format support varies by structure. For example, AVIF and HEIC use a constrained item layout for supported EXIF and XMP handling. HEIC cleanup can return a partial result and leaves codec-level metadata outside its scope.
How to remove supported embedded metadata
In Inspect File or the AI metadata remover, review the report and the selected categories. To remove a PNG prompt or workflow, select Text. Select XMP if it contains records you also want to remove. Check the other categories before starting.
Choose Clean selected. The tool removes whole selected categories where supported. For example, removing PNG text can remove both a workflow and an unrelated caption. Embedded C2PA has a separate category. If provenance matters, keep the original and review how edits affect credential validity on any copy.
Review the outcome and choose Download new copy when offered. AiMetaOff rechecks the processed file and leaves the original intact. A limited result can leave data outside the checked or supported structures.
How to verify the downloaded copy
Open the exact download in Inspect File. Compare the same categories and check for the specific values you intended to remove. Read any warnings or retained-item notes and open the image to confirm that it displays as expected. Some records can remain by design, such as a minimal JPEG orientation record. If you re-export later, check that final file as well. This comparison covers the inspected structures in your copy. Provider history, remote credentials, watermarks and post labels are outside the cleanup scope.
FAQ
Do AI images contain EXIF?
Some do. An exporter or editor can write EXIF fields such as orientation or software. Some AI PNGs instead carry generation settings in PNG text. The image format alone tells you little about which fields are present.
Can a viewer recover a missing prompt or seed?
A viewer can show values it finds in readable records. If those values are absent from the inspected file, check the generator's job history, project or saved workflow. AiMetaOff reads stored values and leaves prompt inference from pixels outside its scope.
Does a Content Credential mean the image is AI-generated?
Look at the assertion. C2PA can document camera capture, editing, AI generation and other actions. If a credential records AI use, read what action it describes and review the validation and trust results before relying on it.
Can an image have both Content Credentials and an invisible watermark?
Yes. OpenAI documents both C2PA and SynthID for supported image outputs. The two signals require different checks, so a report about one leaves the other unresolved.
Can cleaning metadata remove an AI label from a post?
AiMetaOff works on supported records in a local file. The service that hosts the post controls its label, and it can use other signals such as an uploader's disclosure.