SpatialAvatar-0: High-Quality 4D Head Avatar with Multi-Stage Reconstruction

Abstract

High-quality 4D head avatars from one or a few source portraits are central to telepresence, AR/VR, and digital-human interaction. 3D Gaussian Splatting (3DGS) has emerged as the dominant representation, with two complementary regimes (generalizable feed-forward predictors and per-subject refiners) maturing in parallel. However, existing feed-forward predictors are trained on a single dataset family with a hard-coded source count, inheriting the corresponding domain bias. Per-subject refiners require 300K--600K iterations and rely on adaptive densification that destroys upstream Gaussian layouts, preventing the two regimes from sharing a representation end-to-end. To bridge both regimes we propose SpatialAvatar-0 on a shared FLAME-mesh-bound Gaussian representation: a feed-forward generator with a parameter-free K-source mean-pool and a monocular-temporal → multi-view-spatial two-phase schedule that anchors against identity-prior collapse onto the smaller multi-view set. We further introduce a 10K-iter layout-preserving per-subject refinement loop that freezes the FLAME-binding and Gaussian count and replaces densification with a three-component anti-spike regularization. On VFHQ/HDTF cross-domain zero-shot we surpass the in-domain leader GAGAvatar by +1.5 dB PSNR despite never training on either test domain, and on the SplattingAvatar monocular benchmark we lead every reported metric, surpassing the 300K-iter GeoAvatar by +1.3 dB PSNR at up to 60× shorter per-subject schedule than common SOTA baselines.

+1.5 dB

PSNR over GAGAvatar zero-shot

+1.3 dB

PSNR over 300K-iter GeoAvatar

60×

shorter per-subject refinement

Key Contributions

K-Source Feed-Forward Generator

A K-source-variable feed-forward FLAME-mesh-bound Gaussian generator with a monocular → multi-view two-phase training schedule, anchored by L2-SP and a 25% NeRSemble cross-time mix against identity-prior collapse on the smaller multi-view set.

Layout-Preserving Refinement

A 10K-iter layout-preserving per-subject refinement loop with a three-component anti-spike regularization replacing densification, leading every reported metric on the SplattingAvatar leaderboard at up to 60× shorter schedule than common SOTA per-subject baselines.

Comprehensive Evaluation

Comprehensive cross-domain and per-subject experiments on VFHQ, HDTF, and the SplattingAvatar monocular benchmark, with ablations validating every design choice.

Method Overview

We frame head avatar reconstruction as two coupled stages over a shared FLAME-mesh-bound 3D Gaussian representation. Stage 1 (feed-forward): a single network f_θ ingests K ∈ {1,2,3,4} source portrait images and emits face-bound 3D Gaussians in one forward pass. Stage 2 (optional per-subject optimization): starting from f_θ's output for a chosen reference frame, we run 10K iterations of photometric refinement against the target video. Stage 1 is trained in two phases on the same architecture: monocular-temporal pretraining (Phase 1, CelebV-HQ) and multi-view-spatial post-training (Phase 2, NeRSemble); the variable source count K ∈ {1,2,3,4} during training exposes f_θ to monocular and multi-view contexts within a single training distribution.

Demonstrations

In ID

Ours	GAGAvatar	CVTHead	GPAvatar	Portrait4D-v2

Cross ID

Reference	Driving	Ours	GAGAvatar	CVTHead	GPAvatar	Portrait4D-v2

Citation

@misc{wang2026spatialavatar0,
      title={SpatialAvatar-0: High-Quality 4D Head Avatar with Multi-Stage Reconstruction}, 
      author={Yiran Wang and Zeyu Zhang and Yuanming Li and Ziming Wang and Yang Zhao},
      year={2026},
      journal={Tech Report},
}