Apple GHARP: Real-Time Gaussian Head Animation From a Reconstruction Prior

Apple Machine Learning Research has published GHARP: Real-Time Gaussian Head Animation from Large-Scale Reconstruction Prior, a paper for the Asian Conference on Computer Vision (ACCV). The claim in one line: animate a 3D Gaussian head in real time, up to 13 times faster on an NVIDIA A100 than earlier work, using 8 times fewer Gaussians, while reaching state-of-the-art quality on the Ava-256 benchmark. Those numbers are Apple's own, taken from its research page, and the paper is on arXiv.

This is a research result, not a product launch. No code is linked and no device ships with it today. But it sits at the intersection of three things this blog has tracked all month: Apple's steady stream of on-device research, the move of 3D reconstruction from slow optimisation to feed-forward models, and the question of who controls realistic digital faces. Below is what GHARP does, how the two-stage design works, what the page does not say, and what to watch next.

TL;DR: GHARP at a glance

What problem does GHARP solve?

A head avatar has to do two jobs. It must look like one specific person, and it must move when that person's expression changes. Classic neural avatar systems pay for the first job at runtime, which is slow. 3D Gaussian splatting, the technique of representing a scene as many small colored 3D blobs that can be drawn very quickly, made rendering fast. The hard part that remained was driving those Gaussians with expressions without a heavy network and without the artifacts that appear around hair, shoulders and clothing.

GHARP's answer is to split the work. Per Apple's description, an identity stage runs offline and builds a representation of the subject's geometry and appearance, and an animation stage runs at runtime. The offline step can be as expensive as it likes, because the user only waits for it once. The runtime step is a lightweight network that predicts expression-dependent residuals, meaning small corrections added on top of the stored identity.

How does the two-stage design work?

Stage one: the identity, built offline

The identity stage builds on what Apple calls a large-scale reconstruction prior: a pretrained model that already knows how to turn images of a person into a 3D representation. Because that prior lays out space consistently, GHARP can fuse several input views into a compact, fixed-size canonical Gaussian representation. A fixed size matters for engineering: it makes memory and compute predictable, which is what a phone needs.

Stage two: animation inside the latent space

The runtime network does not animate from scratch. It operates in the semantically structured latent space of the pretrained reconstruction model. In that space, an expression change such as opening the mouth or raising a brow stays spatially contained, so the network only has to predict small local residuals instead of re-deriving the whole head. Apple says this containment is what makes the residual prediction efficient.

The body alignment network

The less obvious contribution is a body alignment network. Expression codes describe only the face, so the position of the body and clothing in a target image is underspecified. If you train anyway, the model averages over conflicting body appearances, which shows up as blur and temporal flicker. The alignment network aligns the person's body in the target image with the reference input images, which removes that ambiguity from the training signal. In plain terms: it stops the model from being punished for things the face controls cannot explain.

What are the results, and what is missing?

Apple reports three results: state-of-the-art quality on Ava-256, up to 13x faster on an A100, and 8x fewer Gaussians. Fewer Gaussians is the number that matters most for phones, since memory bandwidth and sorting cost scale with the count.

What the research page does not give, and what readers should look for in the paper itself:

• Absolute frame rates. "13x faster" is relative to baselines; the summary gives no frames-per-second figure and no phone timings.
• Training data. Only Ava-256 is named. How the reconstruction prior was trained is not described on the page.
• Identity cost. The offline stage may be slow, and the page does not say how slow or how many input images are needed.
• Code and weights. None are linked.
• Consent and misuse analysis. The page does not discuss how a realistic, real-time head animator could be abused.

Treat the headline numbers as vendor-reported until independent groups reproduce them. That is normal for a fresh paper, not a criticism.

Why does it matter?

First, real-time avatars are a feature race. Video calls, spatial computing headsets and games all want a believable digital you without a studio capture. A method that stores a compact identity and animates it with a small network is the shape of solution that fits a headset or a phone. Apple does not say this is its goal; it says the runtime stage is designed for mobile devices.

Second, it continues a pattern. Apple's research group has been publishing efficiency-first work across modalities: four-step image generation with normalizing trajectory models, a tri-modal masked diffusion model trained from scratch, and a paper on using idle Mac GPUs as a cloud for distributed inference. GHARP is the vision-side version of the same theme: pay the heavy cost once, keep the runtime tiny.

Third, it shows where 3D research is heading. The wave of feed-forward reconstruction work, such as Meta's GenIA and SAM 3D Objects, LingBot-Map's streaming 3D reconstruction and World Labs' Atlas world model, is producing priors that other systems can build on. GHARP is a concrete example of reusing such a prior as a latent space for animation rather than just for static reconstruction.

How does it relate to Apple's earlier head work?

The research page lists two related Apple papers. HeadsUp, from May 2026 at ECCV, is described as a scalable feed-forward method for reconstructing high-quality 3D Gaussian heads from multi-camera setups. HUGS: Human Gaussian Splats, from December 2023, represents an animatable human together with a scene using 3D Gaussian Splatting. Read together, the line runs from per-scene optimisation (HUGS) to feed-forward reconstruction of heads (HeadsUp) to cheap runtime animation of those heads (GHARP). The page itself does not state this progression; it is explainx.ai's reading of the listed related work.

What should builders and readers do with this?

If you build avatar, telepresence or game tooling, the useful takeaways are architectural:

1. Separate identity from motion. Spend compute offline on the person, keep the per-frame network small.
2. Animate in a latent space that is spatially organised. Local edits in latent space map to local edits on the face.
3. Fix your supervision. If conditioning only describes part of the target (face but not body), align the rest or the model will blur.
4. Wait for code. Nothing here is runnable yet; check the arXiv page and Apple's research page for a release.

For privacy-minded readers, a faster path to a convincing animated face of a real person from a few photos is exactly the capability that deserves safeguards. Apple's page does not address them, so watch for a limitations or ethics section in the paper.

What are people likely to ask next?

Is this the same as Memoji or Persona? The page does not connect GHARP to any shipping Apple feature, and we will not guess. It is a research method evaluated on a benchmark.

Why Gaussians instead of meshes or NeRFs? Gaussians render quickly and can represent fine detail like hair. The trade-off is that you need many of them, which is why cutting the count by 8x is a headline result.

Does a bigger A100 speed-up transfer to a phone? Not automatically. GPU datacenter throughput and mobile memory limits differ. That is why the missing on-device numbers matter.

Bottom line

GHARP is a credible, well-scoped research result: a reconstruction-prior latent space used for cheap head animation, plus a fix for body ambiguity that causes flicker. The speed and efficiency claims come from Apple and a single named benchmark. Until code or independent measurements appear, treat it as promising rather than proven, and as one more sign that Apple's research agenda is about doing more on the device in your hand.

Related reading

• Apple normalizing trajectory models: four-step image generation
• Apple's tri-modal masked diffusion paper
• Apple Minions: idle Mac GPUs as a distributed inference cloud
• Meta GenIA and SAM 3D Objects
• LingBot-Map streaming 3D reconstruction
• World Labs Atlas multimodal world model
• Primary sources: Apple Machine Learning Research page and the arXiv paper

Details reflect Apple's research page as of October 26, 2026. Numbers are the authors' own claims and may be refined in the full paper.