Generators

What each generator leaves inside the file

Every image generator leaves a trail, and they do not all leave the same one. Some sign the file with Content Credentials. Some write a single provenance field. Some write the entire prompt into the file. And a few put a watermark into the picture itself, where no tag reader can see it. This page is a working list of what the common tools are known to write. It changes with each release, so treat it as a starting point and read the actual file on the Metadata tab.

Tools that sign with Content Credentials

A C2PA manifest is a signed record naming the tool and the action. It is the tag most platforms look for first. The C2PA guide has the detail.

• ChatGPT and DALL·E. Downloads carry a manifest naming the generator and the application it was made in.
• Adobe Firefly. Content Credentials are attached by default to generated and AI-edited output, and to edits made with generative features in the wider suite.
• Microsoft Designer and Bing Image Creator. A manifest signed by Microsoft.
• Meta AI. A manifest and IPTC fields, plus an invisible watermark in the pixels.
• Sora and other video generators. A manifest in the MP4 container, and a visible moving watermark drawn into the frames.

Tools that write provenance fields

• Midjourney. IPTC and XMP fields marking the digital source as algorithmically generated, and in newer downloads a manifest as well.
• Google Gemini and Imagen. A SynthID watermark in the pixels on every image, plus provenance tags in some outputs.
• Canva, Figma and similar editors. Software names and, for their generative features, digital source type fields.

Tools that write the prompt

Locally run models and the interfaces built around them are the most talkative. A PNG saved from a local Stable Diffusion or Flux setup commonly carries a text chunk with the full positive and negative prompt, the seed, the sampler, the steps, the model name and often the workflow. It is written so the image can be reproduced, and it stays in the file until something removes it. If you have ever wondered how someone read the prompt off a picture, that is how.

What is in the pixels, not the tags

Two kinds of mark are not metadata and are not removed by cleaning it.

• Invisible pixel watermarks. SynthID from Google, the invisible mark in Meta AI output, and similar systems from other labs. They survive cropping, resizing and screenshots to varying degrees, and only the maker's detector reads them. The SynthID guide says plainly what that means.
• Visible marks. A logo in the corner or a moving watermark across a clip is drawn into the picture.

What survives a save

Tags are fragile by nature. A screenshot has none. A picture re-encoded by a social network has none. A picture opened and re-exported by an editor keeps some and loses others, and often adds its own. Pixel watermarks are the opposite: designed to survive all of that. Which is why a generated image can carry no tags at all and still be identifiable by the generator's own detector, and why a clean report does not prove a photo. The checking guide goes through how to read the evidence.

Removing the tags

Everything in the first three sections is metadata and comes out on the Images tab or the Video tab, without re-encoding. The AI metadata guide covers what that achieves and where it stops.

Common questions

Open the tool