How Google SynthID watermarks text and images, and why decoding is the bottleneck
What happened
Computing Life reported on October 16 how Google SynthID watermarks text and images differently. Text watermarks are embedded through sampling scores as the model generates each word, and detection does not require calling the large model. Diffusion models have no discrete characters to pick from, so image watermarks are added after the whole image is generated by a separate encoder. The report says watermarks have been added to more than 10 billion images and video frames, and SynthID-O can embed up to 136 bits of identity information. Decoder calls may outnumber encoder calls by orders of magnitude, making them the core bottleneck.
Written by AI from the coverage · updated 55 minutes ago
Coverage
Follow the reports to see the story from different sides.
- Computing Life · Share · YageGoogle SynthID 详解：为什么文字边写边打水印，图片却要等画完再补？
Google SynthID 的文字水印在模型逐词生成时通过采样打分嵌入，检测无需调用大模型；图像水印则必须在整张图生成后由独立编码器补打，因为扩散模型没有离散可选字符。图像侧已为超过 100 亿张图像和视频帧打水印，SynthID-O 可嵌入多达 136 bits 身份信息，解码器调用频次可能比编码器高出几个数量级，成为系统核心瓶颈。
Heat over time
Not enough continuous observations to draw a trend yet.