The new cloud-based system was trained using a 15TB library of 4D scan data.
Ziva Dyamics has launched ZRT Face Trainer, their new machine-learning-trained cloud-based facial rigging service meant for games and real-time work.
The new toolkit, available for selected users to test for free, allows transforming "any face mesh into a high-performance real-time puppet in one hour".
The Ziva team states they trained the new cloud-based automated facial rigging platform to recreate a range of expressions of human actors using a 15TB library of 4D scan data. The system is said to convert uploaded character head meshes into a "real-time puppet" that can express over 72,000 facial shapes within an hour.
- The ZRT Face Trainer is built on a library containing over 15TB of 4D scan data.
- ZRT face puppets can express over 72,000 training shapes, as well as novel face poses.
- ZRT face puppets are only 30MB at runtime and run at real-time frame rates (3ms/frame on a single CPU thread).
Please note that in order to process head meshes, they first need to be cleaned and retopologised manually to match a standard point order. They also need to be subdivided for convincing deformations.