How to remove OpenAI SynthID?

Client is running everything I deliver through three detectors and one of them is flagging the images specifically, not the text. I am told this is a watermark baked into the pixels rather than metadata.

Things I have already tried, none of which cleared it:

• Stripping EXIF and all metadata
• Re-encoding as JPEG at quality 70, then 40
• Screenshotting the image and saving the screenshot
• Mild crop and 2% resize

The last one moved the detector's confidence from 98% to 91%, which tells me it survives a lot.

Before anyone asks why: the client's own policy is no AI imagery, and they commissioned AI imagery. I have raised that. I am asking the technical question because I want to understand what the watermark survives, not because I think the policy is my problem to solve with a filter.

It survives all of that by design — it is embedded in the image content at multiple scales, which is why metadata stripping and re-encoding do nothing. Heavy degradation reduces confidence because you are destroying the image, not removing the mark.

That matches what I saw. The crop that moved it to 91% also made the image visibly worse.

The real answer to your client problem is not technical. They have a no-AI policy and an AI brief. One of those has to change, and if you quietly launder the images you own that decision forever.

That is roughly where I have landed, and I said as much in the post. I still want to know what the mark survives.

We hit this on a retainer last quarter. Resolved it by using AI for layout and ideation and shooting or licensing the final assets. Costs more, ends the argument.

Keeping this open because the technical discussion is legitimate and the thread is being honest about the client side. Replies offering a paid removal service will be deleted — that is a marketplace post and I doubt it works anyway.

Worth noting ad platforms are starting to read these marks for disclosure, not takedown. Whatever you do today, assume the detection gets better and your archive does not.

The uncomfortable version: if three detectors are being run on your delivery, the client does not trust you. That is the actual problem in this thread.

Blunt and probably right. We have had one conversation about the brief; this is the second.