Google DeepMind SynthID

Properties

category: watermarking
maturity: established
approach: Embeds imperceptible signals directly into the structure of content (pixel values, audio waveforms, or using tournament sampling for text probability distributions), ensuring the watermark survives compression or cropping.
access: open-source
owns: Watermarks generative AI outputs across text, image, audio, and video modalities natively within the generation process.
distinctFrom: Unlike metadata approaches that can be easily stripped, it modifies the media itself in ways undetectable to humans but visible to algorithmic detectors.
source: DeepMind SynthID
notes: Integrated into Google Gemini and Imagen, and open-sourced as a text-watermarking toolkit via Hugging Face. OpenAI also applies SynthID alongside C2PA to DALL-E 3 and GPT-4o outputs as a dual-layer provenance approach.

category: watermarking
maturity: established
approach: Embeds imperceptible signals directly into the structure of content (pixel values, audio waveforms, or using tournament sampling for text probability distributions), ensuring the watermark survives compression or cropping.
access: open-source
owns: Watermarks generative AI outputs across text, image, audio, and video modalities natively within the generation process.
distinctFrom: Unlike metadata approaches that can be easily stripped, it modifies the media itself in ways undetectable to humans but visible to algorithmic detectors.
source: DeepMind SynthID
notes: Integrated into Google Gemini and Imagen, and open-sourced as a text-watermarking toolkit via Hugging Face. OpenAI also applies SynthID alongside C2PA to DALL-E 3 and GPT-4o outputs as a dual-layer provenance approach.