Dynamic Embedding & Extraction

v0.2a

This past month, our R&D team has released a state-of-the-art AI model that (1) knows where to put watermark data and (2) knows where to look for watermark data.

Original Image // Model View (Human-Friendly) // Watermarked Image

This new model is now incorporated into our watermarks. As a result, our upgraded watermarks are far less noticeable to the human eye. Not only is the absolute PSNR 25–50% lower than before (and well under that of JPEG Q=95), but because the “noise” is added in a smarter fashion, the practical improvements are even greater.

Evaluation of v0.2 (green), v0.1 (blue), and popular libraries (red). Larger circles indicate higher speed.

At the same time, other factors are not compromised. The watermarks are still recoverable under a variety of edits (e.g. below). Server-side encoding and decoding speeds are still in the range of 20–500ms.

Stay tuned for the many Trufo updates to come!