GHARP is a real-time method for animating 3D human heads from a small number of input images and a driving expression signal. The approach splits the problem into two stages: an offline identity stage that builds a representation of the subject’s geometry and appearance, and a runtime animation stage that predicts expression-dependent residuals. Because the identity stage can be computationally expensive while the animation stage is a lightweight network optimized for mobile devices, the design offers a favorable trade-off between fidelity, quality, and runtime.

The method operates in a semantically structured latent space of a pretrained reconstruction model. This prior keeps expression changes spatially contained, which makes residual prediction efficient, and provides a consistent spatial layout so that multiple input views can be fused into a compact, fixed-size canonical Gaussian representation.

A key limitation addressed by the authors is that expression codes only describe the face, leaving body pose and clothing position underspecified. Without additional handling, the animation network faces an ill-posed mapping and averages over conflicting body appearances, producing blur and temporal flicker. GHARP introduces a body alignment network that learns to align the person’s body in the target image with the input reference images, removing this ambiguity from the training signal.

On the Ava-256 benchmark, GHARP achieves state-of-the-art quality while running up to 13× faster on an A100 GPU and using 8× fewer Gaussians. The paper is authored by a team from ETH Zürich and Google, among other institutions.