You strip the metadata, re-upload the image — and the platform still flags it as AI-generated. This confuses a lot of people, and understandably: it looks like the cleaning tool failed. It didn't. You solved one problem and ran into a second, unrelated one.
Two different signals: declared metadata vs. embedded pixel watermark
C2PA is declared information — a manifest that sits alongside the pixel data, naming the tool and timestamp that produced the image. It's a file attachment, conceptually similar to EXIF. Remove it, and nothing about the pixels themselves changes.
SynthID is a different category entirely: an embedded watermark that Google bakes directly into the pixel values when Gemini or Imagen generates an image. It isn't a metadata field you can strip — it's part of the image data itself, invisible to the eye but detectable by Google's own tooling. There is currently no publicly available way to remove it without materially altering the image.
These two signals get lumped together constantly because both result in an "AI-generated" flag somewhere. But they require completely different countermeasures, and only one of them is something a metadata cleaner can touch.
What a metadata cleaner actually removes
A tool like Azaiscale's free metadata cleaner strips every declared layer: C2PA/JUMBF content credentials, EXIF (camera fields, GPS), XMP and IPTC (authorship, software tags). For models that rely solely on this kind of provenance — which is most of them — that's a complete fix. We walk through exactly what each layer contains in our metadata explainer.
What it can't touch
No metadata tool — this one included — can remove a pixel-level watermark. SynthID is the main example in production today, specific to Google's Gemini and Imagen models. Some research-stage watermarking systems from other labs work similarly, though none are deployed as broadly yet. If a platform runs a detector for that specific watermark, a metadata-cleaned image can still be identified as AI-generated, because the signal it's checking was never in the metadata to begin with.
Practical takeaway by model
| Model / tool | What it embeds | Does a metadata cleaner fully solve it? |
|---|---|---|
| ChatGPT / DALL·E | C2PA + EXIF | Yes |
| Midjourney | C2PA + EXIF | Yes |
| Flux | C2PA + EXIF | Yes |
| Google Gemini / Imagen | C2PA + EXIF + SynthID (pixel watermark) | Partially — metadata only, SynthID remains |
In practice: if you're working with ChatGPT, Midjourney or Flux output, cleaning metadata is the whole solution. If you're working with Gemini or Imagen output and a platform specifically checks for SynthID, no metadata tool changes that outcome.
For models that rely on metadata alone, this fully solves it.
Try the free metadata cleanerFrequently asked questions
Is SynthID detectable by the human eye?
No. It's designed to be imperceptible — a statistical pattern in pixel values, not a visible mark or overlay.
Does upscaling or heavy editing remove SynthID?
Not reliably. Google has stated the watermark is designed to survive common transformations like cropping, compression and moderate edits. Extreme transformations may degrade detection confidence but aren't a guaranteed removal method.
Is there any legitimate reason to want an image unmarked?
Yes — privacy and simple workflow hygiene are the most common ones: removing your own generation history from a file before archiving or sharing it doesn't require any dishonest claim about the image's origin.
Skip the ambiguity entirely
If provenance tracking is a recurring headache, the cleanest fix is generating from a pipeline that doesn't introduce it in the first place. Azaiscale, built on Google Nano Banana Pro, delivers photos without EXIF or C2PA markers by default, with a consistent persona across every batch.
Related reading
Azaiscale generates photos that come out already clean — no EXIF, no AI-provenance markers.
Create clean AI photos with Azaiscale