Apple's GHARP animates 3D heads in real time on mobile-ready networks
Apple researchers have presented GHARP, a method that animates 3D human heads in real time from a few input images of a person and a driving expression signal. According to the paper, it reaches state-of-the-art quality on the Ava-256 benchmark while running up to 13× faster on an A100 GPU with 8× fewer Gaussians.
- GHARP animates 3D human heads in real time from a few input images and a driving expression signal.
- An offline identity stage is separated from a lightweight animation stage optimized for mobile devices.
- A body alignment network reduces blur and temporal flicker caused by underspecified body pose and clothing.
- According to the authors, it achieves state-of-the-art quality on Ava-256 while running up to 13× faster on an A100 GPU with 8× fewer Gaussians.
A two-stage approach to head animation
The paper, titled "GHARP: Real-Time Gaussian Head Animation from Large-Scale Reconstruction Prior", was published in October 2026 and is listed for the Asian Conference on Computer Vision (ACCV). It falls under Apple's computer vision research area.
The authors split the problem into two parts. An identity stage builds a representation of the subject's geometry and appearance offline. An animation stage then predicts expression-dependent residuals on top of that representation at runtime.
According to the researchers, this separation offers a favorable trade-off between fidelity, quality and runtime. The identity stage can be expensive because it runs ahead of time, while the animation stage uses a lightweight network that is optimized for mobile devices.
Working in the latent space of a reconstruction model
GHARP performs animation in a semantically structured latent space of a pretrained reconstruction model. The authors say that in this space, expression changes remain spatially contained, which makes predicting the residuals efficient.
The reconstruction prior also provides a consistent spatial layout. This allows several input views to be fused into a compact, fixed-size canonical Gaussian representation of the subject.
Fixing blur and flicker with a body alignment network
The paper points to a problem shared by expression-driven avatar methods. Expression codes describe only the face, so body pose and clothing position are underspecified in the input. According to the authors, the animation network then faces an ill-posed mapping and averages over conflicting body appearances, which produces blur and temporal flicker.
To address this, the team added a body alignment network. It learns to align the person's body in the target image with the input reference images, which removes the ambiguity from the training signal.
Results and context
The authors report state-of-the-art quality on the Ava-256 benchmark. The speed figures are stated for an A100 GPU: up to 13× faster, with 8× fewer Gaussians. The source does not give further runtime numbers or details on specific mobile hardware.
The work is listed alongside related Apple research on 3D Gaussian heads, including HeadsUp, a feed-forward method for reconstructing high-quality 3D Gaussian heads from large-scale multi-camera setups, published in May 2026 for ECCV.
For the field, the results suggest that real-time, high-quality head avatars built from only a few images may become more practical, particularly because the animation network is designed with mobile devices in mind. The source does not mention any product plans.
