Generator comparison
What Each AI Image Generator Puts in Your Files (2026)
AI image generators now mark their output in up to four ways: a signed C2PA manifest, plain metadata labels such as the IPTC digital source type, an invisible watermark in the pixels, and a visible logo. They do not all use the same layers. This table shows what each one documents, and each linked page explains the details and lets you check and clean a file.
Reviewed
Which layers each generator uses
BlankAI can remove the first two columns, because they are file metadata. The last two are part of the image and stay after cleaning.
| Generator | C2PA manifest | Readable metadata (EXIF, XMP, IPTC, PNG text) | Invisible watermark | Visible mark |
|---|---|---|---|---|
| ChatGPT Images | Yes | Sometimes | Yes | No |
| Gemini / Nano Banana | Yes | Sometimes | Yes | Sometimes |
| Midjourney | No | Yes | Not documented | No |
| Meta AI / Muse | Not documented | Sometimes | Yes | Sometimes |
| Grok Imagine | Not documented | Not documented | Not documented | Yes |
| FLUX | Sometimes | Sometimes | Sometimes | No |
| Adobe Firefly | Yes | Yes | Not documented | No |
| Stable Diffusion / ComfyUI | No | Yes | Sometimes | No |
| Microsoft Copilot / Bing | Yes | Not documented | Not documented | Sometimes |
| Dreamina / Seedream | Yes | Sometimes | Yes | Yes |
| Apple Intelligence | Not documented | Yes | Announced | No |
“Not documented” means the vendor has not said either way. “Sometimes” depends on the product, plan or settings. Reviewed September 25, 2026.
Remove Metadata from ChatGPT Images
ChatGPT and OpenAI API images carry a C2PA manifest and, since May 2026, a SynthID watermark. See what each layer holds, remove the metadata, and know what stays.
Read and cleanRemove Metadata from Gemini and Nano Banana Images
Gemini and Nano Banana images carry C2PA credentials, SynthID and sometimes a visible sparkle. See which layers a metadata cleaner removes and which stay.
Read and cleanRemove Metadata from Midjourney Images
Midjourney downloads store your prompt, parameters, Job ID and username in plain metadata, plus an IPTC AI label. Remove them in your browser before sharing.
Read and cleanRemove Metadata from Meta AI Images
Meta AI images use IPTC labels, visible marks on older Imagine output and the invisible Content Seal on Muse Image. See what metadata cleaning can and cannot change.
Read and cleanRemove Metadata from Grok Imagine Images
Grok Imagine images carry a visible Grok watermark that cannot be turned off, and xAI documents no C2PA. See what a metadata cleaner does for Grok files.
Read and cleanRemove Metadata from FLUX Images
FLUX API images carry signed C2PA; local FLUX runs can write your prompt into EXIF and add an invisible watermark. See what each setup embeds and what BlankAI removes.
Read and cleanRemove Content Credentials from Adobe Firefly and Photoshop AI Edits
Adobe Firefly and Photoshop generative edits attach Content Credentials that mark a file as AI-made or AI-edited. See what they record and how to remove them.
Read and cleanRemove Stable Diffusion and ComfyUI Metadata
AUTOMATIC1111, Forge and ComfyUI save prompts, seeds, models and full workflows inside PNG files. Remove that PNG info in your browser and keep transparency.
Read and cleanRemove Metadata from Microsoft Copilot and Bing Image Creator Images
Microsoft Copilot, Designer and Bing Image Creator images carry C2PA Content Credentials, and visible watermarks are now opt-in. See what cleaning removes.
Read and cleanRemove Metadata from Dreamina, Seedream and Jimeng Images
ByteDance's Dreamina, Seedream and Jimeng mark images with visible AI labels, C2PA, invisible watermarks and China's AIGC metadata label. See what cleaning changes.
Read and cleanRemove Metadata from Apple Image Playground and Clean Up Photos
Image Playground and Photos Clean Up write AI labels into image metadata, and Apple plans SynthID for iOS 27. See what is in your file and what cleaning changes.
Read and cleanWhy generators mark their output in 2026
Marking is no longer optional for the large providers. Article 50 of the EU AI Act has applied since 2 August 2026 and requires providers of generative AI to mark output in a machine-readable, detectable way; systems already on the market have until 2 December 2026. The Commission's Code of Practice, signed by Google, OpenAI, Meta, Microsoft, Anthropic, Black Forest Labs and others, asks for at least two layers: signed metadata plus an imperceptible watermark.
California's AI Transparency Act became operative the same day and requires covered providers to embed a latent disclosure in images, video and audio. China has required implicit metadata labels since September 2025, South Korea's AI Basic Act took effect in January 2026, and India's amended IT Rules require labels and permanent identifiers where feasible.
SynthID is becoming the shared watermark
Google's SynthID is now used beyond Google. OpenAI adds it to ChatGPT and API images since May 2026, and Apple has said Apple Intelligence images will carry it after a software update later this year. Google reports more than 100 billion images and videos marked. Meta went its own way with Content Seal, and TikTok and ByteDance use their own watermarks.
The practical result: removing metadata no longer makes an image from these services unrecognizable to the company that made it. Their checkers read the watermark, not the file's metadata.
Why the metadata layer still matters to you
Metadata is where the personal details are. Midjourney writes your username and prompt into the file. Local Stable Diffusion and FLUX tools save prompts, seeds and whole workflows. Phone photos carry GPS coordinates and serial numbers. That is the part a metadata cleaner is for, and it is the part you can check yourself before you share a file.
What a metadata cleaner can and cannot change
| Layer | What it holds | BlankAI |
|---|---|---|
| EXIF and GPS | Camera or phone model, serial numbers, capture time and location coordinates. | Removed |
| XMP and IPTC | Creator tool, edit history, captions, credits and the IPTC Digital Source Type AI label. | Removed |
| C2PA Content Credentials | Who or what created the file and how it was edited, signed by the issuing app. | Removed |
| PNG text chunks | Stable Diffusion prompts, seeds and model names, ComfyUI workflows. | Removed |
| Invisible watermarks | SynthID (Google and OpenAI images), Meta's Content Seal, TikTok and ByteDance watermarks. | Not removed |
| Visible watermarks | Logos, sparkle icons and corner badges drawn onto the image. | Not removed |
| AI classifiers | Models that judge an image by how it looks. File metadata plays no part. | Not affected |
Questions
Which AI image generators add C2PA Content Credentials?
As of September 2026: OpenAI (ChatGPT, Codex and the API), Google (the Gemini app, Vertex AI and Google Ads), Adobe Firefly and Photoshop generative edits, Microsoft Copilot, Designer and Bing Image Creator, Black Forest Labs' FLUX API, and ByteDance's Dreamina. Midjourney and xAI's Grok Imagine do not document C2PA.
Which generators add an invisible watermark?
Google and OpenAI use SynthID. Meta uses Content Seal on Muse Image. ByteDance says Dreamina output carries an invisible watermark. Some local pipelines, such as Stability's SDXL scripts and the FLUX.1 reference code, add a DWT-DCT watermark. Apple has announced SynthID for a later iOS 27 update.
Can BlankAI remove invisible watermarks?
No. BlankAI removes file metadata: EXIF, GPS, XMP, IPTC, C2PA and PNG text chunks. Invisible watermarks are part of the pixels, and BlankAI does not alter pixels to defeat them.
Sources
- Cooley: EU AI Act transparency obligations take effect (August 2026)
- European Commission: Code of Practice on AI-generated content
- Morgan Lewis: California AI disclosure rules become operative (August 2026)
- Google: New ways to identify AI-generated media (May 2026)
- OpenAI: Advancing content provenance (May 2026)
- Bird & Bird: China's measures for labeling AI-generated content