Apple's GHARP animates 3D heads in real time with 8 times fewer Gaussians, and runs up to 13 times faster on an A100
AspireAI community · Apple Machine Learning Research (ACCV)
Apple's Machine Learning Research group published GHARP, a method that animates 3D human heads in real time from a few input images of a subject plus a driving expression signal, accepted at the Asian Conference on Computer Vision. It splits the problem in two: an identity stage builds the subject's geometry and appearance offline, and a lightweight network at runtime predicts expression-dependent residuals on top of that prior, so the expensive half can run once while the frequent half stays cheap.
The animation happens in a semantically structured latent space of a pretrained reconstruction model, where expression changes stay spatially contained and residual prediction stays efficient. That prior also gives a consistent spatial layout, so multiple input views fuse into one compact, fixed-size canonical Gaussian representation. The method adds a body alignment network, because expression codes describe only the face and leave body pose and clothing underspecified - without it the network averages conflicting appearances and produces blur and temporal flicker.
The result is state-of-the-art quality on the Ava-256 benchmark while running up to 13x faster on an A100 GPU with 8x fewer Gaussians. The honest limitation is scope: the target is expression-driven head animation, the reported gains are on a single benchmark, and the underlying paper is the arXiv preprint linked from Apple's page (arXiv:2610.10945).
https://machinelearning.apple.com/research/gharp-head-animation
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.