FastGHA: Generalized Few-Shot 3D Gaussian Head Avatars with Real-Time Animation

How to read this

Category

research paper, feed forward 3D Gaussian head avatar reconstruction and real time facial animation

Contributions

• Proposes a feed forward network that reconstructs a canonical per pixel 3D Gaussian head avatar from only a few input images, with no per identity optimization at inference
• Fuses multi view image features from DINOv3 and a Stable Diffusion VAE encoder through a transformer based encoder to improve reconstruction fidelity
• Extends the explicit Gaussian representation with per Gaussian features driven by a lightweight MLP dynamic network, predicting Gaussian deformations from expression codes to enable real time animation
• Uses point maps from a pretrained large reconstruction model, VGGT, as geometry supervision to improve surface smoothness and robustness
• Reports outperforming prior feed forward approaches such as Avat3r in both rendering quality and animation speed, with reconstruction in under one second

Context

This paper sits in the line of feed forward 3D Gaussian head avatar work that followed Avat3r and Facelift, which themselves built on 3D Gaussian Splatting and neural radiance field avatar reconstruction. Where Avat3r uses cross attention blocks to model facial animation from expression codes, this paper argues that approach is too slow for real time use and proposes the lighter per Gaussian MLP deformation network instead. No builds_on ancestors are listed in the archive dataset, but the paper positions itself explicitly as a response to Avat3r's limitations.

Correctness

The claims rest on quantitative experiments comparing rendering quality and inference speed against Avat3r and other state of the art feed forward baselines, plus ablations for the DINOv3 and VAE feature fusion and the VGGT geometry supervision. As an arXiv preprint it has not been through peer review, and it is trained and evaluated on captured multi view head video datasets, so its behavior outside that data distribution, such as stylized or non photoreal characters, is untested here.

Clarity

Dense but conventional research paper prose in the CVPR style, clear for anyone comfortable with recent Gaussian splatting and NeRF avatar literature, harder going for readers unfamiliar with feed forward reconstruction terminology.

How to read it

Five minute pass: read the abstract and Figure 1 to see the shape of the pipeline, a few images in, an animated avatar out. Second pass: read the introduction's three contribution bullets and skim the two stage pipeline description in Section 3, canonical reconstruction followed by MLP based deformation, to see where the speed comes from. A third full pass matters mainly for practitioners evaluating Gaussian avatar techniques for production, who will want the full architecture and loss details plus the quantitative comparison tables further into the paper.