GHARP enables real-time 3D head animation on mobile devices

  • PublicPublic
  • 26 Oct, 2026
GHARP enables real-time 3D head animation on mobile devices

Apple researchers have introduced GHARP, a novel method for animating 3D human heads in real time using a few input images and a driving expression signal. The core innovation lies in splitting the process into two stages: an offline identity stage that builds a detailed representation of the subject's geometry and appearance, and a runtime animation stage that predicts expression-dependent residuals. This separation allows the identity stage to be computationally expensive, while the animation stage uses a lightweight network optimized for mobile devices. This creates a favorable trade-off between visual fidelity, quality, and runtime performance.

The animation stage operates within a semantically structured latent space derived from a pretrained reconstruction model. This approach ensures that expression changes remain spatially contained, making the prediction of residuals highly efficient. The reconstruction prior provides a consistent spatial layout, which facilitates the fusion of multiple input views into a compact, fixed-size canonical Gaussian representation. By leveraging this structured space, the system can process complex facial movements without requiring heavy computational resources during the rendering phase, maintaining the high frame rates needed for interactive applications.

A significant challenge in expression-driven avatar methods is that expression codes typically describe only the face, omitting body pose and clothing position. This omission leaves these regions underspecified, leading to an ill-posed mapping where the animation network might resort to averaging conflicting body appearances, resulting in visual blur and temporal flicker. To address this, GHARP incorporates a body alignment network that learns to align the person's body in the target image with the input reference images. This alignment removes ambiguity from the training signal, ensuring that non-facial elements remain stable and consistent during animation.

The method achieves state-of-the-art quality on the Ava-256 benchmark while significantly improving efficiency. On an A100 GPU, GHARP runs up to 13 times faster than comparable methods while utilizing 8 times fewer Gaussians. This reduction in computational load and increase in speed makes the technology viable for real-time applications on consumer hardware. The combination of high-quality reconstruction and efficient runtime processing positions GHARP as a practical solution for developers seeking to integrate realistic, real-time head animations into mobile and XR platforms without compromising performance.