Does Cropping or Resizing Remove SynthID? What Actually Works
Cropping and resizing feel like they should erase a watermark, but SynthID is built to survive them. Here is what the testing shows and what actually disrupts the signal.
Cropping does not reliably remove SynthID, and neither does resizing. It is the obvious first thing to try: if a watermark lives in a corner, trimming the corner should kill it. But SynthID has no corner. It is embedded across the entire image, so removing part of the frame leaves most of the signal exactly where it was. This post covers how much survives common edits, why, and what genuinely disrupts the watermark.
Why cropping does not work
SynthID is Google DeepMind’s imperceptible watermark, embedded into pixel and frequency-domain patterns at generation time. Crucially, the signal is distributed redundantly: the same watermark information is repeated across many regions of the image. That design is what lets it survive cropping, resizing, compression, and format conversion far better than a visible logo. When you crop, you delete some copies of the signal but keep plenty of others. If you are new to the concept, our explainer on what SynthID is covers the fundamentals.
How much survives each edit
| Edit | Rough detection survival | Notes |
|---|---|---|
| Minor edits (brightness, small resize) | ~85–95% | Signal is designed for this |
| 25% crop | ~80% | Most redundant copies remain |
| 50%+ crop | Drops sharply | You also lose half the image |
| Lossy re-compression | High | Robust to JPEG-style loss |
| Format conversion | High | Survives JPEG/PNG/WebP changes |
| Heavy multi-step editing | Degrades most | The realistic disruptor |
These are approximate ranges from watermark-robustness testing, not fixed guarantees, and they depend on the image and the detector. The pattern is consistent, though: the edits people reach for first (crop, resize, re-save) are precisely the ones SynthID was engineered to shrug off.
What actually disrupts SynthID
- Full diffusion regeneration. Re-synthesizing the image rewrites pixels everywhere, disturbing the redundant signal rather than trimming copies of it.
- AI re-rendering (img2img). Similar effect, though it can drift further from your original.
- Heavy, multi-step editing. Aggressive successive transforms degrade the signal, usually at the cost of visible quality.
Research such as the UnMarker project reports regeneration-style attacks succeeding around 79% of the time against SynthID; Google disputes those numbers. The honest conclusion either way is that disruption is a probability, not a switch. Delete SynthID uses a local SDXL-based regeneration pass to disturb the embedded signal. Results vary, and it is not a guarantee against every proprietary detector.
Do not confuse cropping with verifying
Even after an edit, the only way to know whether SynthID is still detectable is to check. Detection is effectively Google-only, through the SynthID Detector portal and the Gemini app; there is no open public API, and third-party tools give likelihood scores rather than cryptographic proof. Our guide on how to check an image for SynthID walks through it, and remember that the embedded signal is separate from file metadata and Content Credentials.
Last reviewed July 12, 2026. This guide is general product and publishing information, not legal advice.