Feed-forward Gaussian Registration for Head Avatar Creation and Editing

Malte Prinzler; Paulo Gotardo; Siyu Tang; Timo Bolkart

cs.CVMar 16, 2026

Feed-forward Gaussian Registration for Head Avatar Creation and Editing

Authors: Malte Prinzler, Paulo Gotardo, Siyu Tang, Timo Bolkart

Organizations: ETH Zürich · Google

Abstract

We present MATCH (Multi-view Avatars from Topologically Corresponding Heads), a multi-view Gaussian registration method for high-quality head avatar creation and editing. State-of-the-art multi-view head avatar methods require time-consuming head tracking followed by expensive avatar optimization, often resulting in a total creation time of more than one day. MATCH, in contrast, directly predicts Gaussian splat textures in correspondence from calibrated multi-view images in just 0.5 seconds per frame, without requiring data preprocessing. The learned intra-subject correspondence across frames enables fast creation of personalized head avatars, while correspondence across subjects supports applications such as expression transfer, optimization-free tracking, semantic editing, and identity interpolation. We establish these correspondences end-to-end using a transformer-based model that predicts Gaussian splat textures in the fixed UV layout of a template mesh. To achieve this, we introduce a novel registration-guided attention block, where each UV-map token attends exclusively to image tokens depicting its corresponding mesh region. This design improves efficiency and performance compared to dense cross-view attention. MATCH outperforms existing methods in novel-view synthesis, geometry registration, and head avatar generation, while making avatar creation 10 times faster than the closest competing baseline. The code and model weights are available on the project website.

Figures & tables

Figure 2: Overview . Given calibrated multi-view input images, MATCH first predicts a coarse mesh registration using a pretrained network. We obtain RGB and XYZ textures combined with learnable positional embeddings to encode UV tokens and follow GS-LRM [ 67 ] to tokenize the input images. The image and UV tokens serve as input to a transformer with two alternating attention blocks. In the novel registration-guided attention block, we render UV coordinate images from the input views, and for each UV token restrict the attention to image tokens displaying the relevant mesh region. The subsequent grouped attention block performs attention across the UV tokens and the tokens of each input image separately. The transformer outputs processed UV tokens that are projected into a texture of Gaussians.

Figure 3: Correspondence score estimation between image tokens and UV tokens. To ease visualization, the full mesh is rasterized in overlay with the UV renders and patch sizes are increased.

Figure 4: Novel view synthesis comparison on Ava-256 [ 45 ] . MATCH exhibits superior synthesis quality.

Dataset	Method	LPIPS $\downarrow$	CSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	L1 $\downarrow$	L2 $\downarrow$
Ava-256 [ 45 ]	GPAvatar [ 7 ]	0.303	0.507	21.304	0.714	0.046	0.009
	FastAvatar [ 63 ]	0.285	0.667	15.880	0.730	0.074	0.032
	LAM [ 21 ]	0.273	0.678	15.898	0.734	0.078	0.035
	Avat3r [ 29 ]	0.274	0.626	22.722	0.745	0.039	0.007
	FaceLift [ 44 ]	0.208	0.868	21.661	0.825	0.038	0.010
	Ours	0.163	0.928	23.680	0.848	0.027	0.008

Table 1: Novel view synthesis results on Ava-256 and NeRSemble.

Figure 5: Novel view synthesis on NeRSemble. Ours (Ava) / Ours (NeRSemble) are trained on Ava-256 and NeRSemble only, respectively.

Panel of Figure 6 — Figure 6: Ablation experiment on Ava-256.

	LPIPS $\downarrow$	CSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	L1 $\downarrow$	L2 $\downarrow$
Dense Attention	0.221	0.849	20.364	0.794	0.041	0.013
w/o Sapiens	0.202	0.907	22.104	0.819	0.034	0.010
w/o Skipconn.	0.192	0.913	23.075	0.816	0.032	0.009
Orig. TEMPEH	0.190	0.909	22.775	0.823	0.033	0.009
$H_{\text{uv}}=W_{\text{uv}}=256$	0.194	0.907	22.830	0.817	0.033	0.009
$H_{\text{uv}}=W_{\text{uv}}=512$	0.190	0.914	22.829	0.822	0.032	0.009

Table 2: Ablation experiments on Ava-256.

Figure 7: Interpolation between Gaussian textures by factor $\gamma$ .

Panel of Figure 8 — Figure 8: (Top) Semantic editing: Replacing a source’s nose with that of a target. (Bottom) Expression transfer: Jaw articulation and mouth interior are transferred from a target to a source identity.

Panel of Figure 9 — Figure 9: Geometry reconstruction comparison. The heatmaps visualize the Euclidean distance between the predicted and ground truth vertices. ’TEMPEH Glob. ∗ ’ is the adapted version of TEMPEH’s global stage that we use for the coarse mesh registration.

	Full		Face		Mouth		Eyes
	P2P $\downarrow$	P2S $\downarrow$	P2P $\downarrow$	P2S $\downarrow$	P2P $\downarrow$	P2S $\downarrow$	P2P $\downarrow$	P2S $\downarrow$
TEMPEH Glob. ∗	7.34	2.76	3.84	1.50	2.69	1.15	2.27	1.09
TEMPEH [ 5 ]	7.84	2.18	4.04	0.97	2.99	0.69	2.66	0.62
Ours	6.69	2.10	3.18	0.88	2.14	0.72	1.54	0.61

Table 3: Quantitative geometry reconstruction comparison on Ava-256. Scores are reported in mm. P2P: Point-to-Point distance. P2S: Point-to-Surface distance. ’TEMPEH Glob. ∗ ’ is the adapted version of TEMPEH’s global stage that we use for the coarse mesh registration. ’Face’ is the facial area w/o mouth, eyes, and ears.

Panel of Figure 10 — Figure 10: Qualitative comparison for image-based self-reenactment (top) and cross-reenactment (bottom) of the subject-specific avatars.

	Reconstr.	Self-Reenactment			Cross-Reenactment
	Time $\downarrow$	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	CSIM $\uparrow$	EmoL1 $\downarrow$
GA [ 52 ]	15.5h	0.233	0.755	21.248	0.629	10.618
RGBAvatar [ 32 ]	11.4h	0.233	0.782	22.185	0.657	10.333
GEM [ 70 ]	45.3h	0.214	0.778	21.761	0.800	9.866
Ours	4.6h	0.174	0.809	24.122	0.813	9.837

Table 4: Quantitative comparison of subject-specific head avatars.

Figure 11: Limitations. a) Direct expression transfer ( Section 4.3 ) may incur identity leakage for dissimilar identities. b) The meshes extracted from the predicted Gaussian 3D location textures ( Section 4.4 ) can exhibit self-intersecting surfaces. c) The subject-specific head avatars ( Section 3.2 ) are bound to interpolations of the training expressions and do not track eye movement.

Appendix figures & tables22 assets

Supplementary material from the paper’s appendix.

Appendix

Iteration	Stage Name	Training Datasets	Loss Weights
0 - 100k	Geometry Only	Ava-256 [ 45 ]	$w_{\text{geometry}}=1\times 10^{-3}$ $w_{\text{reg}}=1\times 10^{-3}$ $w_{\text{lpips}}=0$ $w_{\text{L1}}=0$ $w_{\text{SSIM}}=0$
100k - 400k	Geometry & Apparance	Ava-256 [ 45 ]	$w_{\text{geometry}}=1\times 10^{-3}$ $w_{\text{reg}}=1\times 10^{-3}$ $w_{\text{L1}}=0.8$ $w_{\text{SSIM}}=0.2$ $w_{\text{LPIPS}}=0$ until 150k, then linearly increasing to $1.0$ until 200k
400k - 860k	Mixed Training	Ava-256 [ 45 ] & NeRSemble [ 28 ]	$w_{\text{geometry}}=(\text{Ava-256: }1\times 10^{-3};\text{NeRSemble: }0)$ $w_{\text{reg}}=1\times 10^{-3}$ $w_{\text{L1}}=0.8$ $w_{\text{SSIM}}=0.2$ $w_{\text{LPIPS}}=1$

Appendix

Table 5: MATCH loss weights in different training stages.

Parameter	Value
UV texture resolution $H_{\text{uv}}\times W_{\text{uv}}$	$1024\times 1024$
UV token patch size $p_{\text{uv}}$	$16$
Image token patch size $p_{\text{img}}$	$8$
Token dimension $d$	$512$
Number of input images $V$	$12$
Image resolution $H_{\text{img}}\times W_{\text{img}}$	$640\times 512$

Appendix

Table 6: MATCH hyperparameters.

Figure 13: Novel view synthesis comparison against CAP4D on Ava-256.

Panel of Figure 14 — Figure 14: Additional semantic editing results. From top to bottom: Transferring beard and lips, eyes, and hairstyle.

$V_{\text{TEMPEH}}$	$V_{\text{MATCH }}$	LPIPS $\downarrow$	CSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	L1 $\downarrow$	L2 $\downarrow$
12	2	0.213	0.866	20.503	0.795	0.042	0.013
12	4	0.212	0.865	21.078	0.798	0.039	0.012
12	8	0.195	0.907	22.350	0.816	0.034	0.010
12	12	0.187	0.918	23.032	0.825	0.032	0.009
12	16	0.182	0.923	23.367	0.831	0.032	0.009
2	2	0.261	0.689	15.996	0.731	0.071	0.033

Appendix

Table 7: Quantitative ablation of the number of input views to MATCH evaluated on Ava-256. We evaluate two scenarios:

i)

Changing the number of input views to MATCH while keeping the number of inputs to the coarse mesh registration model (TEMPEH) at the default (

V=12

ii)

Changing the number of input views for both TEMPEH and MATCH.

Panel of Figure 15 — Figure 15: Qualitative ablation study for the number of input views to MATCH. We evaluate two scenarios. Top: Changing the number of input views to MATCH while keeping the number of inputs to the coarse mesh registration model (TEMPEH) at the default ( $V=12$ ). Bottom: Changing the number of input views for both TEMPEH and MATCH.

Figure 16: Qualitative ablation study for $k_{\mathcal{T},\text{img}}$ , i.e., the number of image tokens that each UV token attends to in the registration-guided attention blocks. The default value is $k_{\mathcal{T},\text{img}}=100$ .

Panel of Figure 17 — Figure 17: Top: Quantitative ablation study for the number of input views to MATCH on Ava-256. We evaluate two scenarios: $i)$ Changing the number of input views to MATCH while keeping the number of inputs to the coarse mesh registration model (TEMPEH) at the default ( $V=12$ ). $ii)$ Changing the number of input views for both TEMPEH and MATCH. Bottom: Inference speed comparison between our model with the novel registration-guided attention versus a version with dense attention across all UV and image tokens.

Panel of Figure 18 — Figure 18: Robustness to errors in the coarse TEMPEH mesh.

$k_{\mathcal{T},\text{img}}$	LPIPS $\downarrow$	CSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	L1 $\downarrow$	L2 $\downarrow$
25	0.184	0.926	22.908	0.830	0.032	0.009
50	0.183	0.924	23.164	0.830	0.031	0.008
100	0.187	0.918	23.032	0.825	0.032	0.009
150	0.185	0.919	22.951	0.826	0.032	0.009

Appendix

Table 8: Quantitative ablation study for

k_{\mathcal{T},\text{img}}

, i.e., the number of image tokens that each UV token attends to in the registration-guided attention blocks. The evaluations were performed on the Ava-256 dataset. The default value is

k_{\mathcal{T},\text{img}}=100

Panel of Figure 19 — Figure 19: In-the-wild (top) and single-image (bottom) results.

Corresp. Dist. [mm] $\downarrow$	Full Head	Face	Ears	Eyes	Mouth	Scalp
TEMPEH [ 5 ] / Ours	8.9 / 8.0	5.4 / 4.8	10.8 / 9.1	2.5 / 2.1	2.8 / 2.7	13.5 / 12.5

Appendix

Table 9: Quantitative correspondence evaluation.

Figure 20: Additional expression transfer results. Note that we only aim to transfer the oral expression and do not apply any modifications to other regions, e.g., the eyes.

Method	GA [ 52 ]	RGBAvatar [ 32 ]	GEM [ 70 ]	Ours
Stage-Wise Durations	VHAP Tracking: 10.65h Avatar Optimization: 4.83h	VHAP Tracking: 10.65h Avatar Optimization: 0.75h	VHAP Tracking: 10.65h CNN-Avatar Optimization: 27.70h Regressor Training: 6.94h	Coarse Mesh Registration: 0.09h MATCH Inference: 0.44h Canonicalization & PCA Decomp.: 0.16h Emoca & Deca Inference: 0.05h PCA Refinement: 2.75h Expression Regressor Training: 1.14h
Total Reconstruction Time	15.48h	11.40h	45.29h	4.63 h

Appendix

Table 10: Head avatar reconstruction time breakdown. The measurements were conducted on a representative training sequence with 3,212 frames.

Panel of Figure 22 — Figure 22: Qualitative ablation study of the changes applied to GEM [ 70 ] to create subject-specific head avatars from MATCH’s predictions. Ours uses 150 PCA components.

	Self-Reenactment					Cross-Reenactment
	LPIPS $\downarrow$	CSIM $\uparrow$	SSIM $\uparrow$	L1 $\downarrow$	PSNR $\uparrow$	CSIM $\uparrow$	EmoL1 $\downarrow$
Modality-Specific PCAs	0.180	0.879	0.816	0.027	24.376	0.815	9.849
Dynamic Mouth	0.174	0.878	0.809	0.027	24.112	0.811	9.792
# PCA Comp. = 50	0.177	0.876	0.808	0.027	24.126	0.814	9.891
# PCA Comp. = 100	0.175	0.880	0.809	0.027	24.112	0.814	9.907
Ours	0.174	0.880	0.809	0.027	24.122	0.813	9.837

Appendix

Table 11: Quantitative ablation study of the changes applied to GEM [ 70 ] to create subject-specific head avatars from MATCH’s predictions. By default, we use 150 PCA components.

Figure 23: Additional novel view synthesis results on Ava-256 [ 45 ] .

Figure 24: Additional novel view synthesis results on NeRSemble [ 28 ] . Ours (Ava) / Ours (NeRSemble) are trained on Ava-256 [ 45 ] and NeRSemble only, respectively.

Figure 25: Additional interpolation results. $\gamma$ denotes the interpolation factor.

Panel of Figure 26 — Figure 26: Additional self-reenactment results for the personalized head avatars.

Panel of Figure 27 — Figure 27: Additional cross-reenactment results for the personalized head avatars.

3D Avatar Generation

3D avatar generation creates realistic, controllable digital humans from images, video, text, or other inputs.

See all 90 papers

FFAvatar: Feed-Forward 4D Head Avatar Reconstruction from Sparse Portrait Images
Jun 29, 2026Jianjiang Yao, Ke Xian, Renxiang Dai +1Feed-Forward 3D Reconstruction 4D Reconstruction
School of Electronic Information and Communications, Huazhong University of Science and Technology, Wuhan 430074, China
Multi-view Consistent 3D Gaussian Head Avatars 'without' Multi-view Generation
May 24, 2026Aviral Chharia, Fernando De la TorreMulti-View Consistency 3D Gaussian Splatting
Carnegie Mellon University
FFAvatar: Few-Shot, Feed-Forward, and Generalizable Avatar Reconstruction
May 14, 2026Thuan Hoang Nguyen, Jiahao Luo, Yinyu Nie +33D Gaussian Splatting 3D Reconstruction
Snap Inc. · MBZUAI · University of California, Santa Cruz

Jun 29, 2026J/K move · Enter open · S save

FFAvatar: Feed-Forward 4D Head Avatar Reconstruction from Sparse Portrait Images

Jianjiang Yao, Ke Xian, Renxiang Dai, Robert Caiming Qiu

School of Electronic Information and Communications, Huazhong University of Science and Technology, Wuhan 430074, China

Feed-Forward 3D Reconstruction 4D Reconstruction 3D Gaussian Splatting 3D Avatar Generation Sparse-View 3D Reconstruction Human Avatar Animation+2

Open paper AskPDF arXiv

We present FFAvatar, a Transformer-based 3D Gaussian framework for fast construction of high-quality and animatable 4D head avatars from one or more reference portrait images. Unlike existing feed-forward approaches that require a fixed number of input views, FFAvatar supports incremental reconstruction, progressively refining the avatar representation as additional reference images become available. At the core of our method is an alternating attention mechanism that disentangles identity appearance from expression and viewpoint variations, enabling the reconstruction of a canonical 3D appearance that remains consistent across poses and facial expressions. To balance visual fidelity and computational efficiency, we introduce a sparse-to-dense learning paradigm. Coarse appearance features are first learned using sparse primitives anchored to the FLAME vertex level and are subsequently densified in the UV domain to capture fine-grained geometric and texture details. We further propose a plug-and-play motion refinement module that enables subject-specific dynamic personalization by modeling residual motion beyond parametric deformation. Extensive experiments demonstrate that FFAvatar efficiently produces high-fidelity and controllable 4D head avatars, achieving superior flexibility, driving efficiency, and identity-consistent rendering across diverse expressions and viewpoints.

Abstract

3D Avatar Generation

Explore similar work

FFAvatar: Feed-Forward 4D Head Avatar Reconstruction from Sparse Portrait Images

Multi-view Consistent 3D Gaussian Head Avatars 'without' Multi-view Generation

FFAvatar: Few-Shot, Feed-Forward, and Generalizable Avatar Reconstruction