How to Remove a SynthID Watermark From an AI Image

The short answer: you cannot reliably erase a SynthID watermark by cropping a corner or wiping out a logo, because there is no logo to wipe out. SynthID is embedded across the pixels of the whole image, so the only approach that meaningfully disrupts it is regenerating the image data itself. This guide walks through what that means, the exact steps we recommend, and how to verify the outcome honestly. Results vary, and no tool can promise a clean pass against every proprietary detector.

#Why simple edits are not enough

SynthID is Google DeepMind’s imperceptible watermark, woven into pixel and frequency-domain patterns at generation time and distributed redundantly across the frame. That redundancy is the whole point: it lets the signal survive cropping, resizing, compression, and format conversion far better than a visible mark ever could. In practice, detection holds up around 85–95% after minor edits and near 80% even after a 25% crop. To learn what does and does not survive, read our tested breakdown of whether cropping removes SynthID.

The takeaway for removal is simple. Because the signal is spread everywhere, you have to change the pixels everywhere. That is why regeneration (reconstructing the image with a diffusion model) is the approach that actually moves the needle, and why light retouching does not. If you are dealing with a visible mark instead, that is a different problem covered in removing a visible AI watermark.

#What actually disrupts the signal

Regeneration works because it re-synthesizes the image content rather than nudging it. Research such as the UnMarker project has reported success rates around 79% against SynthID using regeneration-style attacks, though Google disputes those figures. Either way, the honest framing is a probability, not a guarantee.

#Step by step with Delete SynthID

1. Confirm your rights. Only process an image you generated, own, or are licensed to edit.
2. Upload the original. Supported formats are JPEG, PNG, WebP, AVIF, HEIC, and HEIF up to 20 MB. Your first image is free.
3. Use the highest-quality source you have. Starting from a re-compressed screenshot only compounds artifacts. Our notes on the best image format for watermark removal explain why.
4. Run the regeneration. The tool reconstructs the image locally with an SDXL-based diffusion pass, disrupting the embedded pattern rather than painting over it.
5. Inspect the result at full size. Check faces, hands, fine text, and geometric edges, which are the areas most likely to drift during regeneration.
6. Verify independently. Do not assume success. Confirm it, as described below.

#How to verify the result

Detection for SynthID is effectively Google-only. There is no open public API; verification runs through Google’s SynthID Detector portal and the Gemini app. Third-party “detectors” return likelihood scores, not cryptographic proof, so treat them as a signal rather than a verdict. For a full walkthrough, see how to check an image for SynthID.

• Run the processed image through Google’s SynthID Detector and compare against the original.
• Check provenance metadata separately at contentcredentials.org/verify, since metadata is a different layer from the embedded signal.
• Remember that a “not detected” result is not a certificate. It reflects one detector at one moment.