RESEARCH · RESEARCH · #2604
GHARP: Real-time Gaussian Head Animation from a Large-Scale Reconstruction Prior
GHARP is a two-stage method that builds an offline identity representation of a subject (geometry and appearance) and a lightweight runtime animation network that predicts expression-dependent residuals in a semantically structured latent space of a pretrained reconstruction model. The paper introduces a body alignment network to resolve under-specified body/clothing regions, reports state-of-the-art results on the Ava-256 benchmark, and claims up to 13× faster runtime on an A100 GPU with 8× fewer Gaussians.
KEY POINTS
- GHARP is a two-stage method that builds an offline identity representation of a subject (geometry and appearance) and a lightweight runtime animation network that predicts expression-dependent residuals in a semantically structured latent space of a pretrained reconstruction model.
- The paper introduces a body alignment network to resolve under-specified body/clothing regions, reports state-of-the-art results on the Ava-256 benchmark, and claims up to 13× faster runtime on an A100 GPU with 8× fewer Gaussians.
- The approach enables higher-quality, low-latency 3D head animation suitable for real-time and mobile applications by separating expensive reconstruction from a lightweight animation stage and reducing compute costs.
WHY IT MATTERS
The approach enables higher-quality, low-latency 3D head animation suitable for real-time and mobile applications by separating expensive reconstruction from a lightweight animation stage and reducing compute costs.