Traditional artist-crafted blendshapes often fail to capture realistic tissue deformations, while physics-based models are too slow for real-time use. A new self-supervised neural approach presented at ACM SIGGRAPH / Eurographics SCA 2025, called NeuRiPhy , tackles this head-on. It learns a neural map from rig controls to deformations that minimize the mechanical energy of an anatomically-based face model. This framework achieves, for the first time, real-time performance of physics-based facial rigs , handling complex non-linear deformations and contact interactions.
A massive breakthrough in recent engine updates is the ability to compress complex, heavy vertex simulations (such as cloth simulation or muscle flexing created in Maya or Houdini) into lightweight, real-time morph targets. The engine trains an ML model on the complex simulation, creating high-fidelity runtime approximations that run smoothly on consumer hardware. 3. High-Density Performance Capture Integration
MorphGS proposes a powerful new paradigm that formulates motion transfer from a monocular video to a 3D character as a target-driven analysis-by-synthesis problem. Instead of a traditional reconstruct-then-retarget pipeline (which can amplify errors), it directly optimizes the target character's morphology and pose using image-space supervision. This results in consistent improvements in contact accuracy and reduced interpenetration, and is applicable to a much wider range of categories than previous methods.