What each AI image generator puts in its files

Four of the seven image generators we could source put a signed C2PA manifest in their files: OpenAI, Google, Adobe and Microsoft. Midjourney and Meta AI write an unsigned IPTC tag instead, and the Stable Diffusion interfaces write something else entirely, your prompt and settings in plain text. The table below is the whole page; the notes under it explain the cells that needed a judgement call.

What each generator writes, as of 23 September 2026

Every cell comes from a page we read, named in the last column and listed under the guide. A dash means we found no statement either way, which is different from “no”.

We left three names out because we could not find a first-party statement for them: Leonardo, Ideogram and Canva. Search results are full of confident claims about all three, mostly on sites selling a remover. None of those pages linked to anything the company itself had published, so we didn’t copy them.

The cells that needed a judgement

Google and C2PA. Google’s November 2025 post names Nano Banana Pro (Gemini 3 Pro Image) on three surfaces. It says nothing about images from the earlier Nano Banana (Gemini 2.5 Flash Image) or about Imagen through other routes, so the cell stops where the post stops. The Gemini and Nano Banana guide covers the sparkle and its setting.

Microsoft’s watermark. MediaPost wrote in October 2023 that Bing Image Creator added Content Credentials along with an “invisible digital watermark”, but its description reads like the C2PA signature itself. The InvisMark paper is real work from Microsoft Responsible AI; the paper does not say it ships in Bing. Nobody offers a public check for it either.

Midjourney is the odd row. IPTC’s May 2023 announcement only says Midjourney would adopt the Digital Source Type marker, with no date. What it actually writes today comes from Numonic, who downloaded PNGs from midjourney.com in March 2026 and found the full prompt with every parameter, the Job ID, a creation time, the IPTC value trainedAlgorithmicMedia, and the account’s username in the Author field. That is one company’s inspection, and Midjourney hasn’t published a spec. We haven’t checked a download ourselves.

Where the table points, if privacy is what you care about

We think the files people should worry about most are the ones with no C2PA at all. Adobe’s own list of what goes into an automatic Firefly manifest is short: a thumbnail, Adobe as issuer, a date, the app, the AI tool, and one of two actions. It is built for other people to verify, and it says little about you. A Midjourney download names your account and repeats your prompt word for word. A ComfyUI PNG carries the whole node graph, down to the file name of any photo you fed into it, as the Stable Diffusion and ComfyUI guide shows from the source code.

Removing any of it is the same operation in the file: drop the text and metadata blocks, leave the pixels. The invisible-watermark column is a separate problem that no metadata tool touches, which is why the guide to the three layers keeps them apart.

Two of the rows have their own pages: ChatGPT images and Firefly and Photoshop.