Watermark Comparison
Watermarking has improved immensely with neural networks; native AI/ML research is needed to stay competitive nowadays. We benchmarked every open-source and commercial watermark we could obtain, using a holistic “practical” evaluation, and PawPrint handily took first place in each benchmark we ran — the fastest to decode, thanks to bespoke compact neural network architectures, with a false-positive rate (on detecting a watermark when none exists) far lower than other watermarking schemes. PawPrint also wins handily on durability and matches the state-of-the-art on imperceptibility, making it especially well-suited for content provenance applications. Details are in the PawPrint model report.
Image Watermarking
pawprint.image leads our 2026 image watermarking benchmark at 94/100, beating pixelseal at 86/100 and the JPEG Trust 2026 winner at 80/100. It is both the fastest and most robust watermark by a comfortable margin. See the image benchmark for more details.
Audio Watermarking
pawprint.audio leads our 2026 audio watermarking benchmark at 96/100, beating audioseal at 72/100; the rest of the field hovers around 60/100. Unlike image watermarking where academic research has been highly active, audio watermarking has not quite caught up to modern neural network capabilities, so the gap between Trufo and other providers is massive. See the audio benchmark for more details.
Video Watermarking
Full report coming soon.
