Apple's GHARP research separates identity reconstruction from live head animation

The October research paper reports a faster Gaussian-avatar method, while its benchmark uses an A100 rather than a shipping mobile product.

GHARP, an Apple-associated computer-vision research project appearing in current news feeds, was first submitted to arXiv on October 7. The paper describes a method for animating a person's three-dimensional head from a small set of input images and an expression signal. It is research accepted to ACCV 2026, rather than a newly announced consumer feature.

The method separates two jobs. An offline identity stage reconstructs the subject's appearance and geometry. A lighter runtime network then predicts expression-dependent changes on top of that representation. The design allows more expensive preparation while keeping the repeated animation work relatively compact.

The authors report state-of-the-art quality on the Ava-256 benchmark and up to thirteen times faster operation on an NVIDIA A100, using eight times fewer Gaussian primitives. Those results describe the paper's evaluated setup. A network designed with mobile efficiency in mind is not the same evidence as a measured shipping-phone implementation.

An interesting part of the work concerns what an expression signal leaves out. Facial expression alone does not specify the body's pose or clothing position. Mixing incompatible examples during training can therefore produce blur or flicker outside the face. GHARP adds a body-alignment network to reduce that ambiguity.

For avatar research, the contribution is a division of computation and a more consistent training representation. It shows how preparation, animation and alignment can be treated separately, making the quality-versus-runtime trade-off easier to study without assuming that a realistic face already constitutes a finished communication product.

Sources

- [GHARP: Real-time Gaussian Head Animation from Large-scale Reconstruction Prior](https://arxiv.org/abs/2610.10945) — GHARP research authors / arXiv