Removed the Watermark but Still Flagged as AI? SynthID vs Visible Marks

You lifted the visible logo, you stripped the metadata — and the image is still detected as AI-generated. Frustrating, but expected: an AI image can carry three separate marks, and the one that’s still tripping the detector is the hardest to remove.

The three layers

• Visible watermark — a logo/badge on the pixels (e.g. the Gemini sparkle). Removed by a visible-mark remover.
• Provenance metadata — C2PA Content Credentials and “Made with AI” tags in the file. Deleted losslessly.
• Invisible SynthID — woven into the pixels. This is almost certainly what’s still flagging your image.

Why removing the first two doesn’t clear the third

The visible logo and the metadata sit on top of or beside the picture, so they come off without changing the image. SynthID is different — it is the pixel data, subtly adjusted. Crop it, compress it, delete every tag: the watermark is designed to survive all of that. So a platform reading SynthID still sees “AI-made.”

What it takes to actually clear it

Only changing the pixels removes SynthID — and doing that without wrecking the image means regenerating it (SDXL-style diffusion) so the embedded signal is gone. The output is a faithful look-alike at full resolution, not a byte-for-byte copy. That’s the trade-off, and it’s why proper SynthID removal runs on a GPU rather than instantly in your browser.

Working with a Google image specifically? See removing the Gemini “Nano Banana” watermark.