GHARP: Real-time Gaussian Head Animation from Large-scale Reconstruction Prior

Ali Benlalah; Sepehr Johari; Patricia Vitoria; Armin Kappeler; Artem Sevastopolsky; Alexander Jung; Gabriele Fanelli; Kevin Mader

cs.CVOct 7, 2026

GHARP: Real-time Gaussian Head Animation from Large-scale Reconstruction Prior

Authors: Ali Benlalah, Sepehr Johari, Patricia Vitoria, Armin Kappeler, Artem Sevastopolsky, Alexander Jung, Gabriele Fanelli, Kevin Mader, +9 more

Organizations: Apple

Abstract

We present GHARP (Real-time Gaussian Head Animation from Large-scale Reconstruction Prior), a method that animates 3D human heads in real time from a few input images of a subject and a driving expression signal. We decouple the problem into an identity stage that builds a representation of the subject's geometry and appearance offline, and an animation stage that predicts expression-dependent residuals on top of it at runtime. This separation offers a favorable trade-off with respect to fidelity, quality and runtime: the identity stage can be expensive while the animation stage runs a lightweight network, optimized for mobile devices. Our method performs animation in a semantically structured latent space of a pretrained reconstruction model, where expression changes remain spatially contained, making residual prediction efficient. This reconstruction prior provides a consistent spatial layout, allowing fusion of multiple input views into a compact, fixed-size canonical Gaussian representation. While this two-stage design improves the runtime-quality trade-off, it still inherits a problem common to all expression-driven avatar methods: expression codes describe only the face and thus omit body pose and clothing position, making these regions underspecified in the input. The animation network faces an ill-posed mapping and resorts to averaging over conflicting body appearances, producing blur and temporal flicker. We address this with a body alignment network that learns to align the person's body in the target image with the input reference images, removing the ambiguity from the training signal. Our method achieves state-of-the-art quality on the Ava-256 benchmark while running up to 13x faster on an A100 GPU with 8x fewer Gaussians.

Figures & tables

Panel of Figure 1 — Figure 1 : Left: GHARP animates a subject from a set of input images and an expression code. Additionally, GHARP enables retargeting of the same expression to new subjects with state-of-the-art quality. Right: runtime on A100 vs. reconstruction quality on Ava-256 for 1, 2 and 4 input images. GHARP is an order of magnitude faster.

Figure 2 : Architecture. Input images are encoded into per-view latents that are aggregated by $F_{C}$ into a fixed-size canonical Gaussian UV map $\mathbf{C}$ and by $F_{D}$ into a deformation latent $\mathbf{D}$ , both independent of the number of input views $N$ . At runtime, only the lightweight animation network operates on $\mathbf{C}$ , $\mathbf{D}$ , and an expression code $\mathbf{w}$ to produce the animated Gaussian UV map.

Figure 3 : Canonical latent blending. The canonical Gaussian UV map consistently combines input view details (e.g. teeth, hair) with high fidelity.

Panel of Figure 4 — Figure 4 : Body Alignment Network (BAN). Top row: BAN takes the body pose from the reference image and the facial expression from the target image, producing a target that is consistent in body appearance with the reference (see cyan boxes). Bottom row: BAN Architecture. Reference and target images are encoded into a shared latent space and blended by predicted weights $\sigma(\mathbf{m})$ and a correction $\boldsymbol{\delta}_{a}$ into an aligned latent $\mathbf{A}_{\ell}$ . Decoding yields the aligned Gaussian UV map $\mathbf{A}$ , which carries the body from the reference and the face from the target. Face- and body-region image losses and an adversarial loss drive the blending (Supp. Mat. 0.A.5 ).

	$\sigma^{t}_{\text{body}}$ $\downarrow$	#G $\downarrow$	LPIPS $\downarrow$	SSIM $\uparrow$	AKD $\downarrow$	CSIM $\uparrow$	PSNR $\uparrow$	ms $\downarrow$
Avat3r	18.8	671K	0.23	0.69	3.5	0.62	21.7	166
FastGHA	22.9	534K	0.23	0.75	3.7	0.51	21.3	26.2
Ours ${}_{\text{+BAN}}$	0.7	65K	0.21	0.74	3.1	0.67	21.9	1.92
Ours	6.0	65K	0.20	0.76	3.1	0.67	22.2	1.92

Table 1 : Results on Ava-256 using 4 input views. #G is the number of Gaussians;

\sigma^{t}_{\text{body}}

measures body stability, the mean per-pixel standard deviation of the body-view render across expressions (0–255 scale, lower is more stable); runtime in ms on A100 GPU. Ours and Ours

{}_{\text{+BAN}}

achieve state-of-the-art results across metrics.

								A100		iPhone 17 Pro
	#I	#G	LPIPS $\downarrow$	SSIM $\uparrow$	AKD $\downarrow$	CSIM $\uparrow$	PSNR $\uparrow$	ms $\downarrow$	FPS $\uparrow$	ms $\downarrow$	FPS $\uparrow$
Avat3r	1	182K	0.26	0.73	9.1	0.44	20.5	43.20	23	–	–
FastGHA	1	150K	0.27	0.76	9.6	0.56	20.3	6.81	147	11.80	85
Ours	1	65K	0.27	0.76	7.9	0.62	20.8	1.92	522	1.45	690
Avat3r	2	365K	0.26	0.74	8.8	0.56	20.8	83.79	12	–	–
FastGHA	2	301K	0.27	0.77	9.3	0.58	20.7	12.97	77	24.36	41

Table 2 : Quality and runtime on Internal10K across varying numbers of input images (#I). #G is the number of Gaussians. Runtime is the animation-stage time; Ours is up to

13\times

faster than FastGHA on the A100 and

31\times

on iPhone 17 Pro Neural Engine (ANE). Pixel-aligned metrics for Ours

{}_{\text{+BAN}}

degrade because the predicted body is aligned with a reference image instead of the ground truth. This does not affect animation (AKD) and identity preservation (CSIM) quality.

Panel of Table 3 — Table 3 : Regional image quality (Ava-256, Internal10K) . We report SSIM and PSNR on GT mask-segmented face and body regions at a fixed cropped frontal camera view. Note that these regional metrics do not directly compare to Tab. 1 or Tab. 2 .

Figure 5 : Temporal stability. The colors show the per-pixel temporal standard deviation across a sequence driven by an expression code See zoomed cyan boxes for jitter unrelated to the expression. Avat3r [ 30 ] , FastGHA [ 22 ] , and Ours all exhibit this jitter, whereas Ours ${}_{\text{+BAN}}$ , trained against body-aligned supervision targets (Fig. 4 ), keeps the body stable. The full video is provided in the Supp. Mat.

Panel of Figure 6 — Figure 6 : Qualitative results with BAN. $\text{Ours}_{\text{+BAN}}$ cleanly recovers fine body details (earrings, tattoos) without the blur or geometric distortion seen in other approaches.

Panel of Figure 7 — Figure 7 : Qualitative comparison on Ava-256. Our method produces sharper facial details (e.g. eyes/teeth/wrinkles) and better expression transfer. FastGHA [ 22 ] tends to produce over-smoothed textures and Avat3r [ 30 ] does not preserve identity as well. The boxes mark the zoom-in region shown as an inset in each panel.

Internal10K Dataset
Configuration	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	AKD $\downarrow$	CSIM $\uparrow$
Full model (Ours)	21.6	0.77	0.24	7.4	0.67
w/o canonical representation	21.2	0.76	0.27	7.5	0.65
w/o passthrough loss	21.3	0.77	0.25	7.5	0.66
w/o deformation latent	21.3	0.75	0.25	9.3	0.59
Full model + BAN	19.8	0.73	0.28	7.5	0.66

Table 4 : Design choices ablation. We individually ablate key components of our pipeline. Each w/o row disables one component while keeping all others at default.

Panel of Figure 8 — Figure 8 : Qualitative evaluation of our design choices. We compare (a) the ground truth (GT) against (b) our full model and ablations removing (c) the canonical representation, (d) the passthrough loss, and (e) the deformation latents. Cyan boxes mark the zoom-in region shown in the bottom-left of each panel.

Figure 9 : Cross-identity retargeting. For each pair, the identity input is shown in the first column, and the expression is driven by the subject in the second column. Right table: Average Expression Distance (AED) metric.

Appendix figures & tables18 assets

Supplementary material from the paper’s appendix.

Appendix

ResNetBlock = GN16 $\to$ SiLU $\to$ Conv $3{\times}3$ , repeated twice, with skip connection.
Type	Channels
Input (concat latents)	$N{\times}(512+6)=2072$
Conv $3{\times}3$	$2072\to 512$
ResNetBlock $\times 8$	$512\to 512$
GN16 $\to$ SiLU $\to$ Conv $1{\times}1$	$512\to 2560$
GN = Group Normalization layer with 16 groups

Appendix

Table 5 : Architecture of

F_{C}

(canonical fusion network). Resolution is constant at

32{\times}32

U-Net Encoder
Type	Channels	Resolution
Input ( $\mathbf{G}_{i}$ or $\mathbf{C}$ )	$d_{G}=23$	$256^{2}$
Conv $2{\times}2$ s2, BN, ReLU	$23\to 64$	$256^{2}\to 128^{2}$
Conv $3{\times}3$ s2	$64\to 128$	$128^{2}\to 64^{2}$
DownBlock	$128\to 128$	$64^{2}\to 32^{2}$
DownBlock	$128\to 256$	$32^{2}\to 16^{2}$

Appendix

Table 6 : Architecture of

F_{D}

(deformation encoder). Per-Gaussian U-Net applied independently to each

\mathbf{G}_{i}

and

\mathbf{C}

ResNetBlock = GN16 $\to$ SiLU $\to$ Conv $3{\times}3$ , repeated twice, with skip connection.
Type	Channels
Input (concat all maps)	$(N{+}1){\times}192=960$
Conv $3{\times}3$	$960\to 512$
ResNetBlock $\times 2$	$512\to 512$
GN16 $\to$ SiLU $\to$ Conv $1{\times}1$	$512\to 96$
GN = Group Normalization layer with 16 groups

Appendix

Table 7 : Architecture of

F_{D}

. Consolidation network applied to the concatenated features from all

N{+}1

maps. Resolution is constant at

64{\times}64

Type	Channels	Resolution
Input $\mathbf{w}$	51	—
MLP (FC $\to$ ReLU $\to$ FC)	$51\to 128$	—
Tile to spatial	128	$64^{2}$
Concat ( $\mathbf{D}$ , tiled $\mathbf{w}$ )	$96{+}128\to 224$	$64^{2}$
Conv $1{\times}1$	$224\to 128$	$64^{2}$
ResBlock $\times 6$	$128\to 128$	$64^{2}$

Appendix

Table 8 : Architecture of

E

(animation network).

Encoder + MLP bottleneck
Type	Channels	Resolution
Input ( $\mathbf{Z}_{i}$ )	$512$	$32^{2}$
DownBlock	$512\to 512$	$32^{2}\to 16^{2}$
DownBlock	$512\to 256$	$16^{2}\to 8^{2}$
DownBlock	$256\to 256$	$8^{2}\to 4^{2}$
Flatten	$4096$	—

Appendix

Table 9 : Architecture of the passthrough network.

	Pretraining	Finetuning
Dataset	Internal10K	Ava-256
Image size ( $H{\times}W$ )	$500{\times}375$	$512{\times}334$
UV map size ( $H_{UV}{\times}W_{UV}$ )	$256{\times}256$	$256{\times}256$
HeadsUp ( $H_{E}$ , $H_{D}$ )	frozen	finetuned on Ava-256
Trained modules	$F_{C}$ , $F_{D}$ , $E$	$F_{C}$ , $F_{D}$ , $E$
Steps	200K	15K

Appendix

Table 10 : Training hyperparameters.

Panel of Figure 10 — Figure 10 : Qualitative comparison between Avat3r implementation in HeadsUp and in our method in a setting without rigging. Main differences in our Avat3r implementation include DINOv2 loss (which explains sharper detail), as well as relaxing feature extractor choice from DUSt3R to DINOv3 (while keeping Sapiens in place), and training with expression conditioning.

Loss	Weight
$\mathcal{L}_{1}$	1.0
$\mathcal{L}_{\text{lpips}}$	10.0
$\mathcal{L}_{\text{adv}}$	0.25

Appendix

Table 11 : BAN loss weights.

	$\sigma^{t}_{\text{body}}$ $\downarrow$	$\Delta E_{76}$ $\downarrow$	TCE $\downarrow$
Avat3r	18.8	10.0	11.7
FastGHA	22.9	12.1	13.4
Ours	6.0	2.8	2.8
Ours ${}_{\text{+BAN}}$	0.7	0.2	0.2

Appendix

Table 12 : Body Temporal Stability on Ava-256. Per-pixel temporal standard deviation of the body-view render in RGB (

\sigma^{t}_{\text{body}}

) and CIELAB (

\Delta E_{76}

), and optical-flow drift (TCE, in pixels), computed across each subject’s expressions; lower is more stable. Ours

{}_{\text{+BAN}}

is the most stable across all three metrics. Best in bold , second best underlined .

Figure 11 : Frozen-body blending vs . BAN. At a mouth opening, a narrow ramp leaves a seam at the neck (top), and a wider ramp changes the chin shape and reduces expression fidelity (bottom). Ours ${}_{\text{+BAN}}$ avoids both.

Device	$N$	ms	FPS	ms	FPS	ms	FPS
		Avat3r [ 30 ]		FastGHA [ 22 ]		Ours
H100	1	21.33	47	2.39	418	1.18	851
	2	41.15	24	4.67	214	1.18	851
	4	78.44	13	9.14	109	1.18	851
iPhone 17 Pro (GPU)	1	–	–	17.6	57	5.1	196
	2	–	–	31.76	31	5.1	196

Appendix

Table 13 : Animation runtime (ms) and FPS. Additional results for animation runtime (ms) and FPS for our method, FastGHA, and Avat3r on different devices across varying numbers of input views (

N

Method	#G	A100	iPhone 17 Pro (ANE)
FastGHA [ 22 ] ( $N{=}1$ )	150K	6.81	11.80
Ours260	260K	3.3	3.4
Ours	65K	1.92	1.45

Appendix

Table 14 : Animation runtime (ms) with

4\times

more Gaussians (Ours260).

	Animation	Rendering	End-to-end
Avat3r [ 30 ]	166.23	1.91	168.14
FastGHA [ 22 ]	26.25	1.53	27.78
LAM † [ 21 ]	-	-	3.56
Ours	1.92	1.33	3.25

Appendix

Table 15 : End-to-end online latency (ms) on an A100 with

N{=}4

input views. Rendering uses gsplat [ 74 ] ; animation times are from Tab. 2 . † Reported by the authors for their 80K-Gaussian model.

Panel of Figure 12 — Figure 12 : Impact of the number of input views. Reconstruction quality scales naturally with the number of input images. Even with $N=1$ , Ours produces sharper reconstruction and preserves identity better compared to prior work.

Method	LPIPS $\downarrow$	SSIM $\uparrow$	AKD $\downarrow$	CSIM $\uparrow$	PSNR $\uparrow$
Ours1k	0.28	0.76	7.8	0.60	20.7
Ours	0.24	0.77	7.4	0.67	21.6

Appendix

Table 16 : Ablation on weaker priors. Impact of reducing the training data for the HeadsUp prior from 10k (Ours) to 1k (Ours1k) subjects.

Panel of Figure 13 — Figure 13 : In-the-wild driving image. The bottom row shows images captured with an iPhone. We use ARKit to extract expression codes from these images and drive the avatar from the input image of the same subject (left column).

Figure 14 : More results for cross-identity retargeting. For each pair, the identity input is shown in the first column, and the expression is driven by the subject in the second column.

Method	#G	LPIPS $\downarrow$	SSIM $\uparrow$	AKD $\downarrow$	CSIM $\uparrow$	PSNR $\uparrow$
UIKA [ 68 ]	134K	0.19	0.79	3.9	0.67	21.3
Ours	65K	0.20	0.76	3.1	0.67	22.2

Appendix

Table 17 : Comparison with UIKA on Ava-256 with 4 input views. Top: reconstruction quality. Bottom: animation runtime. #G is the number of Gaussians. UIKA cannot run on the ANE because of the large LBS matrix multiplications.

Explore similar work

AGORA: Adversarial Generation Of Real-time Animatable 3D Gaussian Head Avatars
Dec 6, 2025Ramazan Fazylov, Sergey Zagoruyko, Aleksandr Parkin +23D Gaussian Splatting 3D Asset Generation
Mohamed bin Zayed University of Artificial Intelligence, Abu Dhabi, UAE · Polynome AI, Dubai, UAE · MTS AI, Moscow, Russia
Split and Drive: Dual-Axis Disentanglement for Real-Time Gaussian Head Avatars
Jul 30, 2026MD Wahiduzzaman Khan, Mingshan Jia, Xiaolin Zhang +23D Gaussian Splatting Disentangled Representation Learning
University of Technology Sydney, Australia · Shandong University of Science and Technology, China
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

Dec 6, 2025, cs.CVJ/K move · Enter open · S save

AGORA: Adversarial Generation Of Real-time Animatable 3D Gaussian Head Avatars

Ramazan Fazylov, Sergey Zagoruyko, Aleksandr Parkin, Stamatis Lefkimmiatis+1

Mohamed bin Zayed University of Artificial Intelligence, Abu Dhabi, UAE · Polynome AI, Dubai, UAE · MTS AI, Moscow, Russia

3D Gaussian Splatting 3D Asset Generation 3D Avatar Generation Human Avatar Animation

Open paper AskPDF arXiv

The generation of high-fidelity, animatable 3D human avatars remains a core challenge in computer graphics and vision, with applications in VR, telepresence, and entertainment. Existing approaches based on implicit representations like NeRFs suffer from slow rendering and dynamic inconsistencies, while 3D Gaussian Splatting (3DGS) methods are typically limited to static head generation, lacking dynamic control. We bridge this gap by introducing AGORA, a novel framework that extends 3DGS within a generative adversarial network to produce animatable avatars. Our key contribution is a lightweight, FLAME-conditioned deformation branch that predicts per-Gaussian residuals, enabling identity-preserving, fine-grained expression control while allowing real-time inference. Identity is further preserved through spatial shape conditioning of the identity branch, and expression fidelity is enforced via a dual-discriminator training scheme leveraging synthetic renderings of the parametric mesh. AGORA generates avatars that are not only visually realistic but also precisely controllable. Quantitatively, we outperform state-of-the-art NeRF-based methods on expression accuracy while rendering at 250 FPS on a single GPU and, notably, at

\sim

9 FPS under CPU-only inference -- to our knowledge the first demonstration of CPU-only animatable 3DGS avatar synthesis. This work represents a significant step toward practical, high-performance digital humans. Project website: https://ramazan793.github.io/AGORA/

Abstract

Explore similar work

AGORA: Adversarial Generation Of Real-time Animatable 3D Gaussian Head Avatars

Split and Drive: Dual-Axis Disentanglement for Real-Time Gaussian Head Avatars

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