ProgressiveAvatars: Progressive Animatable 3D Gaussian Avatars

Kaiwen Song; Jinkai Cui; Juyong Zhang

cs.CVMar 17, 2026

ProgressiveAvatars: Progressive Animatable 3D Gaussian Avatars

Authors: Kaiwen Song, Jinkai Cui, Juyong Zhang

Organizations: University of Science and Technology of China

Abstract

In practical real-time XR and telepresence applications, network and computing resources fluctuate frequently. Therefore, a progressive 3D representation is needed. To this end, we propose ProgressiveAvatars, a progressive avatar representation built on a hierarchy of 3D Gaussians grown by adaptive implicit subdivision on a template mesh. 3D Gaussians are defined in face-local coordinates to remain animatable under varying expressions and head motion across multiple detail levels. The hierarchy expands when screen-space signals indicate a lack of detail, allocating resources to important areas. Leveraging importance ranking, ProgressiveAvatars supports incremental loading and rendering, adding new Gaussians as they arrive while preserving previous content, thus achieving smooth quality improvements across varying bandwidths. ProgressiveAvatars enables progressive delivery and progressive rendering under fluctuating network bandwidth and varying compute and memory resources.

Figures & tables

Figure 2: Overview. We take head video as input then recover a tracked FLAME mesh sequence. We bind 3D Gaussians to the local coordinate frame of each FLAME face. During training, screen-space gradients of the Gaussians drive implicit subdivision of the template mesh across multiple levels, yielding a triangle face forest. At rendering time, we precompute per-face importance score and progressively transmit and render the corresponding Gaussians in decreasing order of importance.

Figure 3: The center row shows the full model containing all 3D Gaussians within one level. Transmitting in descending importance makes early partial renderings closely match the full‑model pixel color because dominant contributors arrive first. In contrast, sending low‑importance Gaussians first re‑normalizes partial weights and amplifies weak contributors, causing noticeable color drift from the full model. This motivates an importance‑first schedule within each level for faithful progressive rendering.

Figure 4: Qualitative results on NeRSemble dataset across different transmission percentages.

	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	#Gaussians	Transmit Data	FPS
	NVS			NES			#Gaussians	Transmit Data	FPS
GaussianAvatars	31.10	0.937	0.064	25.80	0.911	0.076	163,829	41.90 MB	271
5% (Base)	27.89	0.851	0.186	25.13	0.804	0.176	10,144	2.60 MB	291
25%	29.14	0.892	0.080	25.58	0.846	0.124	37,302	9.56 MB	278
50%	30.03	0.904	0.073	25.71	0.884	0.105	84,132	21.56 MB	258
100%	31.47	0.929	0.068	25.89	0.908	0.080	169,438	43.42 MB	260

Table 1: Performance comparison across varying transmission budgets. We report Novel View Synthesis (NVS) and Novel Expression Synthesis (NES) using PSNR/SSIM/LPIPS. We also list the number of Gaussians, the amount of data to transmit (in Megabytes), and rendering speed. Rendering speed is measured on an RTX 4090 at

550\times 802

resolution.

Figure 5: Qualitative comparison with state-of-the-art methods.

Figure 6: Comparison of adaptive and uniform subdivision. Right: visualization of per-face subdivision levels. High-frequency regions like facial hair receive more aggressive splitting, whereas smoother areas require substantially fewer subdivisions.

	NVS			NES
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
PointAvatar	25.8	0.893	0.097	23.4	0.884	0.102
GaussianAvatars	31.1	0.937	0.064	25.8	0.911	0.076
Ours (5%)	27.9	0.851	0.186	25.1	0.804	0.176
Ours (100%)	31.5	0.929	0.068	25.9	0.908	0.080

Table 2: Quantitative comparison with SOTA methods. We denote the best and second best scores in different colors.

Panel of Figure 7 — Figure 7: Progressive streaming and multi-level ablation.

Figure 8: Ablation study on multi-level supervision.

		NVS			NES
Strategy	Budget	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
W/o MLS	35%	20.06	0.812	0.284	19.86	0.798	0.327
W/ MLS	35%	29.87	0.897	0.076	25.65	0.859	0.114
W/ Freeze	100%	31.08	0.925	0.090	25.64	0.901	0.080
W/ MLS	100%	31.47	0.929	0.068	25.89	0.908	0.080
W/ Random	25%	28.40	0.867	0.109	25.46	0.838	0.143

Table 3: Ablation study across all subjects. We report average Novel View Synthesis (NVS) and Novel Expression Synthesis (NES) metrics over all subjects. “W/o MLS” supervises only the finest level, whereas “W/ MLS” supervises all levels.

3D Avatar Generation

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

See all 90 papers

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
Efficient 3D Gaussian Head Avatars for Edge Devices
Oct 7, 2026Umar Farooq, Jean-Yves Guillemaut, Adrian Hilton +13D Avatar Generation 3D Gaussian Splatting
CVSSP, University of Surrey
EvoGS: Constructing Continuous-Layered Gaussian Splatting with Evolution Tree for Scalable 3D Streaming
Jun 5, 2026Yuang Shi, Simone Gasparini, Géraldine Morin +13DGS Compression 3D Gaussian Splatting
National University of Singapore, Singapore · IRIT - University of Toulouse, France · IPAL, IRL2955, Singapore

Dec 6, 2025J/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

3D Avatar Generation

Explore similar work

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

Efficient 3D Gaussian Head Avatars for Edge Devices

EvoGS: Constructing Continuous-Layered Gaussian Splatting with Evolution Tree for Scalable 3D Streaming