EAvatar: Expression-Aware Head Reconstruction

• EAvatar is a 3D head avatar reconstruction framework that enhances fine-grained facial expression modeling through sparse expression control and deformation-aware Gaussian splitting.
• It employs a two-stage pipeline combining structure-aware geometry modeling with dynamic Gaussian refinement, leveraging pretrained generative priors for improved detail and stability.
• Empirical results show superior reconstruction fidelity, visual coherence, and expression control compared to state-of-the-art methods on complex, deformable facial regions.

Searching arXiv for the cited EAvatar and related avatar papers to ground the article in the current literature. EAvatar is a 3D head avatar reconstruction framework built on 3D Gaussian Splatting (3DGS) that is designed to improve fine-grained facial expression modeling, local texture continuity, and reconstruction quality in highly deformable regions such as the mouth, eyes, brows, and teeth. In its specific arXiv formulation, EAvatar denotes the method introduced in “Expression-Aware Head Avatar Reconstruction with Generative Geometry Priors” (Zhang et al., 19 Aug 2025). The framework combines a structure-aware geometry stage with a dynamic Gaussian refinement stage, using a sparse expression control mechanism, deformation-aware Gaussian splitting, and a pretrained generative geometry prior to produce head avatars with improved expression controllability, detail fidelity, and visual coherence while preserving real-time rendering.

1. Definition and problem setting

EAvatar is motivated by a limitation of prior 3DGS-based head avatar systems: although methods such as HeadGaS, GaussianAvatars, SplattingAvatar, and GHA achieve fast rendering and high visual quality, they still struggle with three related issues. First, fine-grained facial expression modeling is weak when deformation is spatially localized. Second, local texture continuity can break in highly deformable regions because Gaussian attributes predicted independently by MLPs become locally inconsistent. Third, geometry is unstable in early training and weak in occluded or underconstrained regions (Zhang et al., 19 Aug 2025).

The framework addresses these issues with two major innovations. One is a sparse expression control mechanism combined with deformation-aware Gaussian splitting, intended to improve local expression modeling. The other is a generative geometry prior from a pretrained large-scale model, used to stabilize and improve early geometry learning. This suggests that EAvatar treats facial motion as locally sparse and structurally constrained, rather than as a uniformly distributed deformation problem.

The method is organized as a two-stage pipeline. Stage I performs structure-aware geometry modeling by learning an implicit signed distance function, extracting a surface with DMTet, and regularizing that geometry with a prior mesh from a pretrained generative model. Stage II performs dynamic Gaussian refinement by building a dynamic Gaussian avatar on top of the geometry, predicting expression- and pose-driven deformation and appearance, applying the controllable Gaussian mechanism, and splitting Gaussians in highly deformable regions (Zhang et al., 19 Aug 2025).

2. Sparse expression control and dynamic Gaussian modeling

A central claim of EAvatar is that only a small subset of Gaussians undergo large motion during facial expression changes. These are treated as “control Gaussians.” The method begins from a canonical neutral Gaussian set

$$\mathcal{G}_0 = \{ \mathbf{X}_0, \mathbf{F}_0, \mathbf{Q}_0, \mathbf{S}_0, \mathbf{A}_0 \},$$

where $\mathbf{X}_0 \in \mathbb{R}^{N \times 3}$ are Gaussian centers, $\mathbf{F}_0$ are feature vectors, $\mathbf{Q}_0$ are rotations, $\mathbf{S}_0$ are scales, and $\mathbf{A}_0$ are opacities (Zhang et al., 19 Aug 2025).

Expression coefficients $\theta \in \mathbb{R}^{d_{\text{exp}}}$ and head pose $\beta \in \mathbb{R}^6$ are estimated from 3DMM fitting. For each Gaussian, the model predicts residual updates with separate expression and pose MLP branches. For position,

$$\mathbf{X}(\theta, \beta) = \mathbf{X}_0 + f_{\text{exp}}^{\text{def}}(\mathbf{X}_0, \theta) + f_{\text{pose}}^{\text{def}}(\mathbf{X}_0, \beta).$$

Analogous residual formulations are used for color, rotation, scale, and opacity. A rigid transform $T(\cdot)$ maps position and rotation into world space, yielding the final expression-aware Gaussian set $\mathbf{X}_0 \in \mathbb{R}^{N \times 3}$0 (Zhang et al., 19 Aug 2025).

Control Gaussians are selected by thresholding the expression-induced displacement magnitude

$\mathbf{X}_0 \in \mathbb{R}^{N \times 3}$1

A Gaussian is selected if

$\mathbf{X}_0 \in \mathbb{R}^{N \times 3}$2

with $\mathbf{X}_0 \in \mathbb{R}^{N \times 3}$3 in experiments. For each control Gaussian $\mathbf{X}_0 \in \mathbb{R}^{N \times 3}$4, neighboring Gaussians are collected within radius $\mathbf{X}_0 \in \mathbb{R}^{N \times 3}$5,

$\mathbf{X}_0 \in \mathbb{R}^{N \times 3}$6

and neighbor motion is propagated by Gaussian-kernel weights:

$\mathbf{X}_0 \in \mathbb{R}^{N \times 3}$7

with

$\mathbf{X}_0 \in \mathbb{R}^{N \times 3}$8

This local propagation is intended to preserve expression coherence around deforming regions and avoid discontinuous motion (Zhang et al., 19 Aug 2025).

The framework further introduces deformation-aware Gaussian splitting. If a Gaussian’s predicted displacement exceeds $\mathbf{X}_0 \in \mathbb{R}^{N \times 3}$9, it is split into two Gaussians initialized near the original location and inheriting its attributes. According to the paper, this increases local representational density in regions undergoing substantial shape change, particularly the mouth, teeth, and eyes (Zhang et al., 19 Aug 2025).

3. Structure-aware geometry modeling and generative priors

The first stage of EAvatar learns an implicit signed distance function

$\mathbf{F}_0$0

where $\mathbf{F}_0$1 is the signed distance and $\mathbf{F}_0$2 is a feature vector. The surface is then extracted via DMTet,

$\mathbf{F}_0$3

yielding a mesh surface from the learned implicit field (Zhang et al., 19 Aug 2025).

The paper argues that early geometry optimization from multi-view images alone is unstable: surfaces can be noisy, contours may be incorrect, occluded regions are underconstrained, and optimization can converge to poor local minima. To counter this, EAvatar uses a high-quality prior mesh from a large-scale pretrained generative model with structured latent representation and a sparse-aware transformer. The prior mesh is denoted

$\mathbf{F}_0$4

It is aligned to the predicted mesh using ICP for coordinate consistency (Zhang et al., 19 Aug 2025).

Instead of imposing vertex-wise supervision, the method applies a global alignment loss based on mesh center and scale. If the prior mesh has center $\mathbf{F}_0$5 and scale $\mathbf{F}_0$6, and the predicted mesh has center $\mathbf{F}_0$7 and scale $\mathbf{F}_0$8, then

$\mathbf{F}_0$9

This loss is used as a global structural guide rather than a dense local target. The paper reports that the prior reduces training time by about 11% and improves geometry quality in contours and occluded regions (Zhang et al., 19 Aug 2025).

This design places EAvatar in a broader trend toward stronger geometric priors in avatar reconstruction. For example, SpatialAvatar-0 also builds on a shared FLAME-mesh-bound Gaussian representation to bridge generalizable feed-forward prediction and per-subject refinement, while preserving Gaussian layout across stages (Wang et al., 14 Jun 2026). EAvatar differs in emphasizing expression sparsity and generative mesh guidance for per-subject head reconstruction.

4. Training objectives, data, and optimization protocol

EAvatar’s Stage I objective is

$\mathbf{Q}_0$0

The reported loss weights are $\mathbf{Q}_0$1, $\mathbf{Q}_0$2, $\mathbf{Q}_0$3, $\mathbf{Q}_0$4, $\mathbf{Q}_0$5, and $\mathbf{Q}_0$6 (Zhang et al., 19 Aug 2025).

Stage II uses a simpler appearance-focused objective,

$\mathbf{Q}_0$7

with $\mathbf{Q}_0$8 and $\mathbf{Q}_0$9. This stage optimizes final appearance quality and perceptual realism, especially on local regions (Zhang et al., 19 Aug 2025).

Experiments are conducted on NeRSemble, a multi-view head video dataset with 16 camera views per subject, multiple expression sequences, “FREE” sequences used for evaluation, and remaining sequences used for training. Preprocessing includes background removal via segmentation, 68 facial landmark extraction, and BFM fitting to estimate 3D landmarks, expression coefficients, and head pose. Evaluation uses PSNR, SSIM, and LPIPS (Zhang et al., 19 Aug 2025).

The reported optimization schedule uses Adam. Stage I runs for 10,000 iterations with batch size 4 and learning rate $\mathbf{S}_0$0. Stage II runs for 500,000 iterations with batch size 1. In Stage II, learning rates are $\mathbf{S}_0$1 for color, deformation, and attribute MLPs, $\mathbf{S}_0$2 for neutral positions and feature vectors, $\mathbf{S}_0$3 for rotation, and $\mathbf{S}_0$4 for scale. Runtime is about 2.5 days per subject on RTX 3090, with rendering speed about 32 FPS (Zhang et al., 19 Aug 2025).

5. Empirical performance and ablation behavior

On self-reenactment, EAvatar is compared with NeRFace, HAvatar, and GHA. The reported averages are: NeRFace at 20.85 PSNR / 0.805 SSIM / 0.248 LPIPS; HAvatar at 21.93 / 0.836 / 0.228; GHA at 24.56 / 0.853 / 0.175; and EAvatar at 25.07 / 0.857 / 0.168 (Zhang et al., 19 Aug 2025). The paper describes these gains as especially strong in reconstruction fidelity and perceptual quality.

Qualitatively, EAvatar is reported to better reconstruct teeth, eye closure, hair detail, and subtle facial expressions. For cross-identity reenactment, the paper reports clearer and more realistic transfer, with better expression transfer, more stable identity preservation across frames, and better local details in the mouth and eyebrows. On held-out novel views, EAvatar reports PSNR 23.22, SSIM 0.882, and LPIPS 0.165, indicating improved 3D consistency under viewpoint changes (Zhang et al., 19 Aug 2025).

The ablation results are central to the paper’s argument. Removing the generative prior and mesh alignment yields worse contours and less stable early geometry. Removing control-point propagation produces blurry or distorted mouths and eyes under strong expressions. Adding the controllable Gaussian mechanism improves local continuity and expression fidelity. Deformation-aware splitting improves teeth and inner-mouth appearance, local geometric resolution, and fine-scale detail in high-motion areas (Zhang et al., 19 Aug 2025).

Threshold selection is also studied. The paper reports that control threshold 0.3 works best overall: values that are too low over-propagate motion, while values that are too high miss subtle expression changes. For splitting, 0.2 provides a good balance; smaller values add too many Gaussians with little perceptual gain, while larger values delay refinement (Zhang et al., 19 Aug 2025). This suggests that the method’s sparsity assumptions are operational rather than merely descriptive.

6. Position within avatar research and adjacent concerns

EAvatar belongs to a rapidly expanding family of avatar systems that differ in embodiment level, controllability, and representation. Within high-fidelity head modeling, HQ3DAvatar uses a fully implicit canonical radiance field with multiresolution hash encoding and a canonical-space optical-flow correspondence loss, and is trained from multi-view data while being driven at test time by monocular RGB video (Teotia et al., 2023). By contrast, EAvatar is explicitly 3DGS-based and centers its contribution on sparse local deformation control and pretrained geometric guidance.

More broadly, avatar research spans full-body generative meshes, editable neural humans, egocentric telepresence, open-domain 4D avatarization, and socially situated self-avatars. GETAvatar generates explicit textured 3D meshes for animatable full-body human avatars and uses normal-map supervision from 3D scans to improve geometry detail (Zhang et al., 2023). NECA decomposes geometry, albedo, shadow, and lighting for granular customization from monocular or sparse-view videos (Xiao et al., 2024). AvatarGo focuses on plug-and-play self-avatars in VR and reports that exact user-specific tracker-to-joint offsets significantly improve Sense of Embodiment relative to fixed offsets (Ponton et al., 2022). These systems address different strata of the same problem space: realism, control, deployment simplicity, and user perception.

The perceptual dimension is particularly relevant to any system that increases realism. “Agent vs. Avatar: Comparing Embodied Conversational Agents Concerning Characteristics of the Uncanny Valley” reports that increased perceived humanness correlates with increased perceived eeriness, with $\mathbf{S}_0$5 and $\mathbf{S}_0$6, and that the more human-like avatars in the study were perceived as both more human and more eerie (Thaler et al., 2021). A plausible implication is that improvements in expression fidelity and local realism, including those pursued by EAvatar, are not purely geometric questions; they also interact with acceptance, affect, and interface design.

The paper on EAvatar does not present a formal limitations section, but practical constraints are evident from the method design. The framework still depends on 3DMM fitting for expression and pose control. The control and split thresholds are fixed empirically and may require tuning across datasets. The method is trained per subject rather than as a universal one-shot avatar model. Although the generative prior improves geometry, the training process still requires substantial optimization time (Zhang et al., 19 Aug 2025).

In this sense, EAvatar is best understood as a specialized head-avatar reconstruction method that refines the 3DGS paradigm along three axes at once: local expression sparsity, adaptive local representational density, and stronger geometry priors. Its significance lies less in replacing the broader avatar literature than in making a precise intervention in a known failure mode of high-fidelity head avatars: the mismatch between globally efficient rendering and locally complex facial deformation.