Apple GHARP: body alignment beats packing more Gaussians for real-time head avatars

Real-time head avatars from a few images still fail in a boring place: faces track, while shoulders, neckline and clothing smear. Apple Machine Learning Research’s October 2026 paper GHARP (Real-time Gaussian Head Animation from Large-scale Reconstruction Prior) treats that smear as a specification problem, not a missing Gaussian budget.

The thesis is sharp. Expression codes describe the face and omit body pose and clothing position. The animation network then faces an ill-posed map and averages conflicting body appearances—producing blur and flicker that more primitives will not fix.

GHARP decouples an expensive identity stage that builds subject geometry and appearance offline from a runtime animation stage that predicts expression-dependent residuals. Animation runs in a semantically structured latent space of a pretrained reconstruction model, where expression changes stay spatially contained. The reconstruction prior also gives a consistent layout for fusing multiple views into a compact, fixed-size canonical Gaussian representation.

The body-alignment network learns to align the person’s body in the target image with the input reference images, removing ambiguity from the training signal before the residual animator learns. The animation network is described as lightweight and optimized for mobile devices; identity work can stay heavy offline.

Reported results: state-of-the-art quality on the Ava-256 benchmark, up to 13× faster on an A100, with 8× fewer Gaussians. That pairing matters for the thesis—if quality rises while Gaussian count falls and runtime improves, the gain is unlikely to be “just denser representation.”

Authors span Apple ML research contributors (Benlalah, Johari, Vitoria, Kappeler, Sevastopolsky and others). The public page is a paper abstract-level disclosure; method sketches beyond the abstract were not fully expanded on the landing page fetched for this column.

The steelman is that better expression/pose representations, multi-view capture, or temporal priors could reduce underspecification without a dedicated alignment network—and Ava-256 SOTA can still hide domain shift to in-the-wild phone selfies, occlusions, or extreme poses. Another counter: two-stage systems add identity enrollment cost that single-stage on-device methods may refuse.

Also thin without the full paper PDF in hand: exact baselines for the 13×/8× claims, failure cases of the alignment net, licensing/data details for the large-scale reconstruction prior, and whether clothes dynamics beyond alignment remain unmodeled. Confidence in the mechanism is high relative to the abstract; confidence in transferable SOTA numbers is medium until tables are checked.

For on-device avatars, the product implication is architectural: spend offline compute on identity priors and keep runtime residual predictors small—but invest modeling capacity in making the training signal well-posed. Packing more Gaussians into a single-stage model is the seductive wrong lever when the labels themselves are ambiguous.

For the broader Gaussian avatar literature, GHARP is a reminder that “real-time” quality regressions often live outside the face crop.

Watch follow-up work that ablates body alignment against denser single-stage Gaussians under matched runtime, and mobile deployments that keep the 13×-class advantage off the A100. If alignment wins those ablations, expression-only driving codes will look incomplete by default. If not, GHARP’s two-stage elegance may still lose to simpler pipelines that cheat with more capture.