GHARP animates a 3D head in real time: 13x faster on A100

Researchers have presented GHARP, a method that animates a person's 3D head in real time from a few photos and a facial expression signal. On the Ava-256 benchmark it delivers state-of-the-art quality, while running up to 13 times faster on an A100 GPU and using 8 times fewer Gaussians. The network that handles animation at runtime is lightweight and optimized for mobile devices.

Two stages instead of one

The central idea is to split the task. The first stage, identity, builds a representation of a specific person's geometry and appearance. It runs offline, so it can be as expensive as needed. The second stage runs at runtime and predicts expression-dependent residual changes on top of the finished representation.

This trade-off between accuracy, quality, and speed is familiar from other tasks: the heavy work moves to where time is unlimited, and the user's device gets the minimum amount of computation. For avatars this matters especially, because animation has to run at a frame rate suitable for video calls or games.

The animation happens neither in pixel space nor in a raw parameter set, but in the semantically structured latent space of a pretrained reconstruction model. There, expression changes stay spatially local, which makes the residuals easier to predict. The same reconstruction prior also provides a single spatial layout, which makes it possible to merge several input views into a compact canonical representation built on a fixed-size set of Gaussians.

Body and clothing: the problem that expression codes are silent about

The two-stage design improves the balance between speed and quality, but it does not remove a problem common to all expression-driven avatars. An expression code describes only the face. Body pose and clothing position are absent from it, so these regions remain undefined in the training signal. The animation network faces an ill-posed task and averages over conflicting variants of the body's appearance. The result is blur and temporal flicker.

The authors address this with a separate body alignment network. It learns to align the person's body in the target image with the input reference photos. The ambiguity disappears from the training signal, and the animation network is no longer forced to average.

Key figures

• Input: several images of a person and a driving expression signal.
• Benchmark: Ava-256, state-of-the-art quality.
• Speed: up to 13 times faster on A100.
• Compactness: 8 times fewer Gaussians.
• Runtime network: lightweight, optimized for mobile devices.
• Identity: offline stage that allows expensive computation.

Where this is useful and what to watch

The practical gain lies in areas where an avatar has to live on the user's device: telepresence, video calls with a 3D representation, games, AR, and VR. Using fewer Gaussians lowers memory and rendering bandwidth requirements, and moving the expensive step offline removes the need to compute identity on every launch.

The limitations are also visible. The abstract gives no data on specific mobile chips, frame rates on a phone, or measurement conditions, so the 13x speedup claim should be checked against the full paper and the code, if the authors publish it. In addition, the offline stage still requires several good-quality photos of the person, which means this method does not cover the "one selfie on the fly" scenario.

For teams building products with digital doubles, GHARP is useful as an architecture reference: separate identity from animation, work in the latent space of a reconstruction model, and explicitly resolve body ambiguity at training time.