Apple presents GHARP method for real-time 3D head animation

Apple researchers presented GHARP, a two-stage method for animating a 3D human head from a few calibrated images and a driving facial-expression signal. The system builds and caches a subject-specific identity representation offline, leaving a lightweight network to predict expression-driven changes at runtime. That division makes the per-frame animation cost independent of the number of views used to construct the subject.

The offline stage fuses features from the input views in a UV-aligned latent space and decodes them into a fixed-size Gaussian UV map anchored to a neutral head template. Each of the roughly 65,000 Gaussians stores position, scale, rotation, opacity and degree-1 spherical-harmonic color coefficients. At runtime, the network takes that cached representation and a 51-dimensional expression code derived from ARKit, then predicts residual updates to the canonical Gaussians instead of reconstructing the identity for every frame.

On the Ava-256 benchmark, the authors trained on 244 subjects and evaluated on the remaining 12, sampling about 1,000 frames per subject. With four input views, they report an average-keypoint-distance score of 3.1, cosine identity similarity of 0.67, PSNR of 22.2, SSIM of 0.76 and LPIPS of 0.20. These results are author-reported and have not been independently replicated.

Apple and the GHARP paper say the animation stage is up to 13 times faster than FastGHA on an Nvidia A100 while using up to eight times fewer Gaussians. In the paper’s Internal10K runtime table, the authors report 1.45 milliseconds, or 690 frames per second, on an iPhone 17 Pro Neural Engine and 1.92 milliseconds, or 522 frames per second, on an A100. Those figures cover the animation stage after expression-independent work is precomputed. Including rendering with four input views, the paper reports 3.25 milliseconds on an A100, compared with 27.78 milliseconds for the authors’ FastGHA implementation and 168.14 milliseconds for their Avat3r implementation.

The comparison does not use official implementations. The paper says public code was unavailable for Avat3r and FastGHA, so the authors used their own reimplementations. Their FastGHA setup also omitted VGGT geometry supervision because they considered precomputing it at the required scale infeasible.

GHARP also addresses a failure mode that arises because facial-expression signals do not specify body pose or clothing position. Training against conflicting body appearances can make a model average them, producing lost detail, blur and temporal flicker. The researchers separately train a Body Alignment Network that combines the target expression with body appearance from a reference image in latent space, then use its aligned rendering as supervision for the animation network. On Ava-256 with four views, they report that this reduced their body temporal-stability metric from 6.0 to 0.7, where lower is more stable.

The paper’s abstract page lists the paper as accepted to ACCV 2026. The authors say GHARP still depends on calibrated camera poses, meaning pose-estimation errors can propagate through the encoder. They identify robustness to temporal variation or noisy inputs and portrait relighting as future work.