How SynthID Works: Frequency-Domain Embedding
Traditional watermarking embeds a visible stamp or appends text tags to the file header. If someone crops the stamp or strips the header, traditional watermarks are lost.
SynthID takes an entirely different approach. Developed by researchers at Google DeepMind, it uses two neural network models:
- The Embedding Model: During image or video generation, it introduces subtle, imperceptible modifications into the frequency domain of the visual data. These changes are balanced so that human eyes perceive zero difference in sharpness, contrast, or color.
- The Detection Model: A specialized detector scans the mathematical relationships across the frequency spectrum. It calculates a statistical confidence score indicating whether Google's model produced the media.
Why SynthID Survives Standard Edits
Google designed SynthID specifically to withstand common post-processing operations:
- Cropping: Because the signal is distributed across the entire canvas rather than isolated in one corner, cropping an image leaves the signal intact in the remaining sections.
- JPEG Compression: The frequency shifts survive lossy quantization down to moderate quality levels.
- Color Adjustments: Routine tweaks to saturation, brightness, and contrast do not destroy the underlying pattern.
The Reality of SynthID Removal
Can any software remove SynthID? The honest answer is no—not without severely corrupting the image.
Because SynthID is woven into the primary frequencies of the image data, the only way to disrupt the signal is through aggressive destructive filters: extreme gaussian blurring, heavy downsampling, or adding overwhelming noise. Doing so destroys the quality of the artwork.
However, what creators can and often do want to remove is the visible corner star. The four-pointed sparkle icon in the bottom-right corner of Gemini images is a separate visible overlay that can be cleanly removed with our Gemini Watermark Remover without altering the rest of your picture.