Dose NOT REMOVE SynthID remove-ai-watermarks.

[MiroriBlue]

Hello Team,

I downloaded it from GitHub and ran it on my local machine, but it detected my SynthID.

I also paid for and tried the version from
https://raiw.cc
,but it still detected my SynthID.

Is this program really working properly?

[wiltodelta]

Sorry it is not working for you, and thanks for trying raiw.cc. Let me help get to the bottom of this.

Two things would help a lot:

1. How are you checking for SynthID? There is no public SynthID decoder, so the only valid check is the vendor's own oracle: the Gemini app "Verify with SynthID" for Google images, or openai.com/verify for OpenAI images. A generic AI-vs-real detector (Sightengine, Hive, and similar) is not a SynthID check and will keep flagging the image even after the watermark is gone. So the first thing I want to confirm is what tool reported "SynthID detected".

2. Could you email me the original image? I cannot reproduce or fix anything without the actual sample. Send it to kuznetsov.va@gmail.com. If you include the email you paid with, I will issue a refund for raiw.cc.

On the local run: the SynthID (invisible) pass only runs when the [gpu] extra is installed. Without it the tool still removes the visible mark and metadata and writes an output file, but the SynthID pixel pass is silently skipped, so the watermark stays. Install the extra and rerun:

pip install 'remove-ai-watermarks[gpu]'
remove-ai-watermarks all image.png -o clean.png

[Patrickickyourass]

I tried the project on pics generated by doubao and Gemini, my purpose was to remove the visiable watermark so it means I don't have to use the gpu version, but neither works, I tried the force mode but it only left a blur on the bottom right corner. Does it have anthing to do with the size of the pic, I just downloaded the original pic file, do I need to conpress the file?

[wiltodelta]

@Patrickickyourass the blur in the bottom-right corner comes from forcing removal without detection: that path runs the Gemini sparkle remover, which inpaints a corner even when there is no sparkle there. For a Doubao or Gemini image, run the default (no force), it auto-detects the mark:

remove-ai-watermarks visible your.png -o clean.png

That finds the Doubao "豆包AI生成" strip, the Jimeng "★ 即梦AI" wordmark, or the Gemini sparkle, and removes what is actually there. No GPU needed for the visible pass.

Size is not the cause and you do not need to compress, visible removal runs at native resolution. Worth upgrading first, v0.13.0 just shipped with better visible removal:

pip install -U remove-ai-watermarks

If it still leaves a mark, email the original image to kuznetsov.va@gmail.com and I will check what is happening.

[Patrickickyourass]

Thanks for replying my comment. And Yes, for some reason, the regular command  remove-ai-watermarks visible could not remove the visible watermark, and I noticed that recently  that Google seems has changed the way they watermark the images, the watermarks are little bit bigger, lightter and moved a little bit to the north-west direction of the images. And i took a few days for your teply to find me, so I had to develop an algorithm of my own, which I had made it a web, to remove just the visible watermark of Gemini-AI images: The Lost Star - AI Image Artifact Purifier Thanks again for developing such a good project, I will follow the progress and use it on to see if there is any new issue and leave my comment.

[wiltodelta]

Thanks for the detailed report, and the north-west / lighter observation is genuinely useful. That moved, fainter sparkle is exactly the case the last two releases went after.

v0.14.0 added a white-core rescue that recovers faint and re-rendered Gemini sparkles by default, and v0.14.1 (just out) is the latest. For screenshots or re-saves where the mark shifted, there is also a sensitivity switch that trusts a borderline sparkle instead of leaving it:

pip install -U remove-ai-watermarks
remove-ai-watermarks visible your.png -o clean.png --sensitivity assume-ai

That assume-ai path is what recovers the ones the strict default still misses.

If it still leaves the mark, email the original to kuznetsov.va@gmail.com and I will check it against the actual sample. And nice that you built your own purifier, glad the problem got solved either way.