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

Yingdong Hu; Yisheng He; Yiming Jiang; Zehong Lin; Steven Hoi; Jun Zhang

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

NeurIPS 2026 Poster

Yingdong Hu^*1, Yisheng He^*†2, Yiming Jiang², Zehong Lin³, Steven Hoi², Jun Zhang¹

¹ HKUST² Alibaba Token Hub, Alibaba Group³ Lingnan University

* Equal Contribution † Corresponding Author

Paper arXiv Video BibTeX

FA-LAM creates an animatable Gaussian full head from a single image in 0.4644 seconds on an H20 GPU, with high-quality novel views and expressive animation.

One-shot animatable full-head reconstruction with fine facial detail and strong large-view synthesis.
Semantic-symmetric attention regularization and a disentangled reconstruction–animation training pipeline.
Multi-view and streaming 4D reconstruction with visibility-aware fusion and constant-size recurrent state.

Abstract

We propose FA-LAM, a Focus-Aware Large Avatar Model for one-shot animatable Gaussian head creation, while simultaneously enabling static 3D and dynamic 4D full-head recovery. The core of our method lies in a thorough analysis of the attention mechanisms and the entangled reconstruction and animation training pipeline adopted by prior state-of-the-art approaches. Our analysis identifies two main factors that compromise the quality of 3D full-head generation: (1) incorrect and noisy attention activations, and (2) conflicts between the tasks of reconstruction and animation. To address the first issue, we introduce a symmetric and semantic attention regularization strategy that leverages the inherent semantics and structural symmetry of human heads. To disentangle the objectives of reconstruction and animation, we develop a novel dual-phase training pipeline that separates the model’s capabilities for large-view hallucination and animation into distinct modules. Moreover, we enhance our model to support multi-view and streaming 4D reconstruction in an efficient and memory-friendly manner through a core autoregressive modification with tailored visibility-aware token fusion. Collectively, these innovations enable FA-LAM to reconstruct animatable Gaussian full heads with superior quality, particularly in fine facial regions and large viewing angles.

Pipeline

Semantic-Symmetric Attention Regularization

We use FLAME-based semantic correspondences and the bilateral symmetry of human heads to guide cross-attention. A Gaussian mixture target and KL-divergence regularization encourage each UV token to focus on its corresponding image region or its visible symmetric counterpart, rather than unrelated facial regions or the background.

Disentangling Reconstruction and Animation

We first learn a cross-view-consistent, canonical full-head representation from multi-view data. We then learn expression-dependent Gaussian residuals, with an additional DPT head for the eye and mouth regions. This separates identity-specific geometry from non-rigid expression changes while preserving large-view reconstruction quality.

Visibility-Gated Autoregressive Fusion

A fixed-size recurrent state accumulates observations over time. Previously visible tokens are retained when currently occluded; newly visible tokens use the current observation; tokens visible in both frames are fused by a lightweight MLP. The recurrent state memory remains constant with respect to the number of input frames.

Attention and gradient analysis

Noisy attention in visible regions and semantically inconsistent attention in occluded regions — Without explicit regularization, attention can become noisy or focus on semantically unrelated regions, particularly for occluded tokens.

Gradient cosine similarity distributions and t-SNE features for reconstruction and animation objectives — 47.7% of gradient pairs have cosine similarity below −0.1, 43.2% are approximately orthogonal, and only 9.1% are positively aligned.

One-Shot Animatable Gaussian Heads

Self-Reenactment and Large-View Rendering

From a single portrait, FA-LAM generates a full-head Gaussian avatar that can be animated and rendered from large viewing angles. The reconstruction preserves identity and fine appearance details without test-time optimization.

Self-reenactment and large-view rendering · 00:24–00:40 of the full demo.

Cross-Identity Reenactment

Motion from a different identity can drive the reconstructed avatar while preserving the appearance of the source portrait. FA-LAM also generalizes to generated portraits with complex makeup and tattoos, and to driving signals from real-world videos.

Qualitative comparison of self-reenactment and cross-identity animation on VFHQ — Qualitative comparisons on VFHQ. FA-LAM improves appearance fidelity, identity preservation, and expression transfer.

Novel-View Synthesis and Multi-Image Reconstruction

FA-LAM produces consistent novel views on NeRSemble-v2 and benefits from additional observations on Ava256. Visibility-gated autoregressive fusion incorporates complementary views without attending to the entire input history.

Qualitative novel-view and multi-image comparisons on NeRSemble-v2 and Ava256 — Novel-view synthesis on NeRSemble-v2 and avatar reconstruction on Ava256.

Streaming 4D Full-Head Reconstruction

With only the reconstruction head enabled, FA-LAM reconstructs temporally aligned Gaussian full heads from monocular video streams. Visibility-aware token fusion preserves reliable observations and progressively fills previously unseen regions.

Monocular streaming 4D reconstruction · 01:32–01:48 of the full demo.

Temporal consistency comparison with visibility-gated autoregressive feature fusion — Visibility-gated fusion improves the temporal consistency of reconstructed regions that are not visible in every frame.

Quantitative Results

We evaluate one-shot avatar generation, novel-view synthesis, multi-image reconstruction, and streaming 4D reconstruction on VFHQ, NeRSemble-v2, and Ava256.

Novel-View Synthesis on NeRSemble-v2

FA-LAM achieves the best performance across all seven reported metrics, with a PSNR of 25.96 dB, improving by 2.50 dB over GAGAvatar (23.46 dB).

Novel-view synthesis on NeRSemble-v2. Source: Table 3, p. 8 of the author-provided NeurIPS manuscript.
Method	PSNR ↑	SSIM ↑	LPIPS ↓	CSIM ↑	AED ↓	APD ↓	AKD ↓
GPAvatar	21.71	0.871	0.319	0.543	0.539	0.541	12.537
Real3DPortrait	20.71	0.885	0.317	0.586	0.511	0.573	11.749
Portrait4D-v2	21.98	0.854	0.301	0.554	0.534	0.610	37.738
GAGAvatar	23.46	0.906	0.157	0.693	0.433	0.217	7.874
LAM	18.32	0.809	0.131	0.538	0.319	0.301	31.479
FA-LAM (Ours)	25.96	0.954	0.087	0.766	0.102	0.128	7.034

One-Shot Avatar Generation on VFHQ

FA-LAM achieves the best reported PSNR, SSIM, LPIPS, CSIM, and APD for self-reenactment, and the best scores across all three cross-identity metrics. LAM retains the best self-reenactment AED and AKD.

VFHQ: seven self-reenactment metrics followed by three cross-identity metrics. Source: Table 2, p. 7 of the author-provided NeurIPS manuscript.
Method	Self PSNR ↑	Self SSIM ↑	Self LPIPS ↓	Self CSIM ↑	Self AED ↓	Self APD ↓	Self AKD ↓	Cross CSIM ↑	Cross AED ↓	Cross APD ↓
GPAvatar	21.04	0.807	0.150	0.772	0.132	0.189	4.226	0.564	0.255	0.328
Real3DPortrait	20.88	0.780	0.154	0.801	0.150	0.268	5.971	0.663	0.296	0.411
Portrait4D-v2	21.34	0.791	0.144	0.803	0.117	0.187	3.749	0.656	0.268	0.273
GAGAvatar	21.83	0.818	0.122	0.816	0.111	0.135	3.349	0.633	0.253	0.247
LAM	22.65	0.829	0.109	0.822	0.102	0.134	2.059	0.651	0.250	0.356
FA-LAM (Ours)	22.91	0.857	0.096	0.844	0.117	0.112	2.464	0.692	0.229	0.234

Single- and Few-Image Avatar Generation on Ava256

Moving from one input image to four increases FA-LAM PSNR from 22.18 to 22.46 dB (+0.28 dB). In the four-image setting, FA-LAM leads all seven reported metrics. In the one-image setting, it leads PSNR, LPIPS, APD, and AKD, but not every metric.

One-Shot (1 Input Image)

Ava256 one-shot avatar creation. Source: Table 4, p. 8 of the author-provided NeurIPS manuscript.
Method	Type	PSNR ↑	SSIM ↑	LPIPS ↓	CSIM ↑	AED ↓	APD ↓	AKD ↓
Portrait4D-v2	FFD	18.58	0.738	0.186	0.598	0.172	0.391	15.672
GAGAvatar	FFD	19.86	0.688	0.254	0.506	0.157	0.428	19.875
LAM	FFD	18.45	0.621	0.289	0.463	0.192	0.557	16.338
InvertAvatar	GAN	19.72	0.695	0.228	0.397	0.189	0.491	12.731
GPAvatar	FFD	19.89	0.699	0.277	0.386	0.201	0.530	24.941
DiffusionRig	DIF	20.71	0.718	0.268	0.389	0.223	0.445	9.360
FastAvatar	FFD	16.92	0.536	0.349	0.379	0.205	0.569	20.842
FA-LAM (Ours)	FFD	22.18	0.733	0.156	0.517	0.171	0.380	6.141

Few-Shot (4 Input Images)

Ava256 few-shot avatar creation. Source: Table 4, p. 8 of the author-provided NeurIPS manuscript.
Method	Type	PSNR ↑	SSIM ↑	LPIPS ↓	CSIM ↑	AED ↓	APD ↓	AKD ↓
InvertAvatar	GAN	20.73	0.724	0.213	0.441	0.183	0.555	12.403
GPAvatar	FFD	18.91	0.675	0.294	0.349	0.204	0.509	22.499
DiffusionRig	DIF	22.09	0.756	0.251	0.487	0.198	0.401	10.431
FastAvatar	FFD	16.97	0.530	0.332	0.446	0.209	0.596	20.911
FA-LAM (Ours)	FFD	22.46	0.766	0.120	0.655	0.153	0.316	4.183

FFD: feed-forward network. GAN: GAN inversion. DIF: diffusion.

Streaming 4D Reconstruction on NeRSemble-v2

Using only monocular input, FA-LAM reconstructs temporally aligned full heads and improves PSNR by 6.01 dB over the four-view Forge4D baseline in this evaluation.

Streaming 4D full-head reconstruction on NeRSemble-v2. Source: Table 5, p. 8 of the author-provided NeurIPS manuscript.
Method	Temporal Alignment	Input Views	PSNR ↑	SSIM ↑	LPIPS ↓
Forge4D	Yes	4	19.32	0.712	0.146
PanoLAM	No	1	20.17	0.917	0.214
LAM	No	1	19.96	0.799	0.145
FA-LAM (Ours)	Yes	1	25.33	0.942	0.096

↑ Higher is better. ↓ Lower is better. Highlighted rows identify our method, not necessarily the best value in every column. Results follow the author-provided NeurIPS manuscript; table values take precedence over inconsistent prose. Comparisons are made within each task and evaluation protocol.

Ablation Studies

Semantic-Symmetric Attention

Full semantic-symmetric regularization improves PSNR from 25.30 without regularization to 25.80, while producing more localized attention and clearer facial details.

Rendering comparison across attention regularization strategies — Qualitative comparison of attention regularization strategies.

Two-Stage Training

Training reconstruction before animation achieves a PSNR of 25.80, compared with 24.99 for joint training. The disentangled strategy preserves detail under challenging novel views.

Efficient and Consistent Streaming

For 20 input frames, visibility-gated fusion uses only 20 MB more memory than frame-by-frame reconstruction, an increase of approximately 0.23%. It improves PSNR by 0.87 dB and reduces TL-STD from 8.141 to 7.253. Full attention runs out of memory in this experiment.

Full ablation results

Attention regularization

Attention regularization ablation; configuration labels follow Table 7. Source: Table 7, p. 9 of the author-provided NeurIPS manuscript.
Configuration	PSNR ↑	SSIM ↑	LPIPS ↓
Direct feature unwarping	23.34	0.871	0.157
w.o. attn-reg	25.30	0.934	0.101
w. sym. attn-reg	25.68	0.932	0.096
w. full attn-reg (Ours)	25.80	0.949	0.091

Attention activation maps with different regularization strategies — Attention maps under different regularization strategies.

Training strategy

Training strategy ablation. Source: Table 6, p. 9 of the author-provided NeurIPS manuscript.
Strategy	PSNR ↑	SSIM ↑	LPIPS ↓
Joint training	24.99	0.938	0.103
Animation only	24.93	0.940	0.111
Reconstruction only	25.65	0.939	0.099
Reconstruction, then animation (Ours)	25.80	0.949	0.091

Autoregressive fusion with 20 input frames

Streaming reconstruction and memory usage. Source: Table 8, p. 9 of the author-provided NeurIPS manuscript.
Configuration	TL-STD ↓	TI ↓	PSNR ↑	SSIM ↑	LPIPS ↓	GPU Memory ↓
Full attention	N/A	N/A	N/A	N/A	N/A	OOM
Framewise reconstruction	8.141	13.731	24.46	0.940	0.106	8751 MB
Autoregressive (no token fusion)	8.253	13.760	24.53	0.940	0.099	8755 MB
Autoregressive (visibility-aware fusion, Ours)	7.253	12.805	25.33	0.942	0.096	8771 MB

Visibility-Based vs. Learnable Gating

Explicit visibility-based gating outperforms learnable gating across all five reported quality and temporal metrics.

Additional ablation of the fusion module. Source: Table 10, p. 17 of the author-provided NeurIPS manuscript.
Configuration	TL-STD ↓	TI ↓	PSNR ↑	SSIM ↑	LPIPS ↓
No Fusion	8.253	13.760	24.53	0.940	0.099
Learnable Gated Fusion	8.317	13.233	24.10	0.939	0.099
FA-LAM (Ours)	7.253	12.805	25.33	0.942	0.096

Computational Efficiency

On an RTX A6000, FA-LAM achieves 243.9 FPS for animation and rendering, with 495.9 ms reconstruction time. Reconstruction is approximately 5.61× faster than LAM-20k (2780.7 ms), while using more GPU memory (6.675 GB vs. 3.613 GB). These measurements are separate from the 0.4644-second H20 inference result reported in Section 4.2.

Computational efficiency on RTX A6000. Source: Table 9, p. 17 of the author-provided NeurIPS manuscript.
Method	Animation & Rendering (FPS) ↑	Reconstruction (ms) ↓	FLOPs (G)	GPU Memory (GB)
FA-LAM (Ours)	243.9	495.9	1701	6.675
LAM-20k	250.1	2780.7	7667	3.613
GAGAvatar	47.6	85.2	570	2.885
Portrait4D-v2	8.1	244.5	838	3.261
InvertAvatar	4.3	164.2	122	4.605
GPAvatar	4.8	255.6	443	5.028

MV-VFHQ Dataset

To provide stronger supervision for large viewing angles, we extend VFHQ with multi-view data. For each frame, Qwen-Image-Edit generates four additional views at ±45° and ±90°. An optimization-based FLAME tracker, initialized with EMOCA, annotates shape, expression, and camera parameters.

BibTeX

@misc{hu2026falam,
  title         = {FA-LAM: Focus-Aware Large Avatar Model for One-Shot 4D Animatable Gaussian Head},
  author        = {Yingdong Hu and Yisheng He and Yiming Jiang and Zehong Lin and Steven Hoi and Jun Zhang},
  year          = {2026},
  eprint        = {2607.20922},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2607.20922}
}

arXiv preprint citation.