Research map / 3D vision / Face and shape

FA-LAM: Focus-Aware Large Avatar Model for One-Shot 4D Animatable Gaussian Head

NeurIPS 2026 · 3D vision › Face and shape · approach: Gaussian and implicit head avatars · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes FA-LAM, a feed-forward large avatar model that adds symmetric semantic attention regularization, a dual-phase reconstruct-then-animate training pipeline, and visibility-gated autoregressive fusion to produce animatable Gaussian full heads from one image or streams.

Problem

Feed-forward one-shot animatable Gaussian head avatars suffer from noisy cross-attention on invisible regions, blurry fine details, and large-view degradation from conflicting reconstruction and animation objectives; naive multi-view/streaming extensions scale quadratically in memory.

Approach

Regularizes UV-token cross-attention toward semantically and bilaterally symmetric image locations via a KL loss against Gaussian mixtures, separates static reconstruction training from expression animation fine-tuning (with residual Gaussians), and fuses streaming UV tokens with a visibility-gated update for constant memory.

Key results

• Single-image cross-reenactment on VFHQ: CSIM 0.692 vs 0.663 best baseline (Real3DPortrait), APD 0.234 best.
• NeRSemble-v2 self-reenactment: PSNR 25.80 and LPIPS 0.091 vs GAGAvatar 23.38 and 0.164.
• 4D streaming on NeRSemble-v2: PSNR 25.33 with one input view, vs Forge4D 19.32 with four views; autoregressive fusion adds only ~20 MB GPU memory where full attention runs OOM.

Surprising findings

• Joint reconstruction and animation training degrades quality; attention-layer gradients show negative cosine similarity for nearly half of task pairs.
• Self-reenactment on VFHQ shows FA-LAM does not win on all metrics: AED 0.117 is worse than LAM's 0.102 and GAGAvatar's 0.111.

Limitations

• Quality at occiputs (back of head) remains weak, as noted by the authors.
• Relies on a coarse FLAME-based UV extraction and FLAME tracking; the MV-VFHQ annotations come from an optimization-based tracker initialized by EMOCA.
• Multi-view training data is partly synthesized by Qwen-Image-Edit, which may introduce view artifacts.

Experimental setup

Datasets: VFHQ NeRSemble-v2 Ava256 MV-VFHQ

Benchmarks: VFHQ NeRSemble-v2 Ava256

Metrics: PSNR SSIM LPIPS CSIM AED APD AKD TL-STD TI GPU memory

Baselines: GPAvatar Real3DPortrait Portrait4D-v2 GAGAvatar LAM InvertAvatar DiffusionRig FastAvatar Forge4D PanoLAM

Builds on: LAM PanoLAM PCGrad DINO

Closest papers by meaning

• GAF: Gaussian Avatar Reconstruction from Monocular Videos via Multi-view Diffusion (CVPR 2025 · similarity 0.89)
• Generalizable and Animatable Gaussian Head Avatar (NeurIPS 2024 · similarity 0.89)
• Gaussian Head Avatar: Ultra High-fidelity Head Avatar via Dynamic Gaussians (CVPR 2024 · similarity 0.88)
• FATE: Full-head Gaussian Avatar with Textural Editing from Monocular Video (CVPR 2025 · similarity 0.88)
• CAP4D: Creating Animatable 4D Portrait Avatars with Morphable Multi-View Diffusion Models (CVPR 2025 · similarity 0.88)
• Synthetic Prior for Few-Shot Drivable Head Avatar Inversion (CVPR 2025 · similarity 0.87)