🧑GHARP Animates 3D Human Heads in Real Time on Mobile Devices

Photorealistic avatars just got cheap enough for your phone

Real-time photorealistic 3D heads just moved from the lab toward your phone. GHARP takes a few input photos plus a driving expression signal and animates a full 3D head live. The trick is splitting the work into two stages: an offline identity stage that builds geometry and appearance, and a runtime animation stage that only predicts expression-dependent residuals. If you're building video calls, telepresence, or avatar features, this is the architecture to study. The animation stage runs a lightweight network built for mobile hardware, so you don't need a GPU server per user. It works in the latent space of a pretrained reconstruction model, which gives a consistent spatial layout and lets multiple input views fuse into one compact canonical Gaussian representation. The numbers: state-of-the-art quality on the Ava-256 benchmark, up to 13x faster on an A100 GPU, and 8x fewer Gaussians. A body alignment network matches the target body to the reference images and strips ambiguity from the training signal. Fewer Gaussians means less memory and less render cost, which is where most avatar pipelines fall over.