SVG-Head: Hybrid Editable Gaussian Head Avatar

• The paper introduces a hybrid model that binds explicit texture images with FLAME-aligned surface Gaussians to enable real-time, localized appearance editing.
• It leverages a dedicated volumetric branch to capture complex regions like lips and hair, ensuring high-fidelity reconstruction even in challenging areas.
• The approach decouples geometry from appearance, overcoming prior limitations of implicit representations in neural head avatars for editable and photorealistic outcomes.

Searching arXiv for SVG-Head and closely related head-avatar Gaussian papers. Surface-Volumetric Gaussian Head Avatar (SVG-Head) is a hybrid head-avatar representation that explicitly models geometry with 3D Gaussians bound on a FLAME mesh and leverages disentangled texture images to capture global appearance, while additional volumetric Gaussians model regions that are difficult to explain with surface-constrained texture sampling alone, especially lips and hair. It was introduced to address a specific limitation of prior neural head avatars: photorealistic rendering and plausible animation had become common, but real-time, localized appearance editing remained difficult because geometry and appearance were encoded implicitly or entangled across Gaussian attributes. SVG-Head therefore separates editable surface appearance from residual volumetric effects, while retaining Gaussian-splatting-based rendering and FLAME-driven animation (Sun et al., 13 Aug 2025).

1. Research problem and representational premise

SVG-Head is motivated by the observation that NeRF-based head avatars encode appearance implicitly in a radiance field, which is powerful for rendering but poorly suited to direct manipulation such as painting makeup, recoloring a facial region, or editing localized texture details interactively. Standard Gaussian head avatars improve rendering speed and animation, but they typically attach color directly to individual Gaussians; geometry and appearance remain entangled, so editing requires optimization over many primitives or external image-based guidance rather than direct texture-space control. The method is framed as a response to this representational bottleneck: explicit editability requires a representation in which appearance can be modified independently of the Gaussian geometry that supports animation (Sun et al., 13 Aug 2025).

The core premise is hybridization. SVG-Head does not reduce the head to a textured mesh, nor does it treat the avatar as a free Gaussian cloud. Instead, it combines a surface-aligned branch that carries globally editable appearance and a volumetric branch that compensates for appearance and geometry not well captured by that surface model. This division of labor is central: surface Gaussians provide correspondence, UV-space editability, and sharp texture images, while volumetric Gaussians preserve reconstruction quality in non-Lambertian or geometrically complex regions.

A common misconception is to treat SVG-Head as merely another FLAME-rigged 3DGS avatar. That characterization misses its main distinction. The paper’s contribution is not only FLAME-bound Gaussian animation, but the introduction of explicit texture images for Gaussian head avatars and the claim that these textures support real-time appearance editing without re-optimization. A second misconception is the opposite one: SVG-Head is not a pure texture-mapped surface model. The volumetric branch remains essential for residual phenomena that do not fit the surface-only assumption.

2. Hybrid FLAME-bound Gaussian representation

SVG-Head uses FLAME as its animating scaffold. As in GaussianAvatars, Gaussians are bound to local frames of FLAME mesh triangles so that pose and expression changes can be transferred through the mesh. For each Gaussian $\mathcal{G}_i$ attached to triangle $j$, the method defines local Gaussian parameters relative to the triangle frame: local rotation $\mathbf{R}_i^l \in \mathbb{R}^{3\times 3}$, local center $\boldsymbol{\mu}_i^l \in \mathbb{R}^3$, and local anisotropic scale $\mathbf{s}_i^l \in \mathbb{R}^3$. The triangle itself has a local frame with rotation $\mathbf{R}_j^t$ and translation $\boldsymbol{\mu}_j^t$, and the global Gaussian parameters are recovered by

$$\boldsymbol \mu_i = k_j \mathbf R^t_j \boldsymbol \mu^l_i+\boldsymbol \mu^t_j,$$

$\mathbf R_i = \mathbf R^t_j\mathbf R^l_i,$

$\mathbf s_i = k_j\mathbf s^l_i,$

where $j$0 is a scalar proportional to the shape of the $j$1-th face. The FLAME parameters are denoted $j$2 for shape, $j$3 for expression, and $j$4 for pose, and are optimized in training (Sun et al., 13 Aug 2025).

The representation contains two Gaussian types. Surface Gaussians, or surf-GS, are constrained to live on the FLAME mesh and derive their color from explicit learnable texture images. These are the editable component. Volumetric Gaussians, or vol-GS, are added around the mesh and store their own appearance directly; they compensate for reconstruction errors in difficult regions after the surface branch has established a sharp, UV-aligned appearance model.

The asymmetry between the two branches is deliberate. Surf-GS are the principal carrier of correspondence and appearance semantics. They are initialized one per triangle at triangle centers and are later densified and split. Vol-GS are residual by design: they preserve animatability by sharing the same local-to-global FLAME binding mechanism, but they are permitted to optimize their own geometric parameters and full color representation.

3. Texture decomposition and mesh-aware Gaussian UV mapping

The surface branch replaces per-Gaussian color storage with an explicit texture-space model. SVG-Head decomposes texture into a static diffuse texture $j$5 and an expression-dependent dynamic texture $j$6. The diffuse texture models the base appearance, including skin tone and makeup-like details. The dynamic texture accounts for expression-correlated changes such as wrinkles, and is generated by convolutional networks conditioned on FLAME expression parameters $j$7. Surface Gaussians retain only high-degree spherical harmonics residuals $j$8 for view-dependent components, while the main view-independent appearance is sampled from the texture images (Sun et al., 13 Aug 2025).

A key technical problem is establishing a stable correspondence between a 3D Gaussian and a point in FLAME UV space. SVG-Head introduces a mesh-aware Gaussian UV mapping that leverages the known UV coordinates of the underlying FLAME mesh. For a ray $j$9 intersecting a Gaussian $\mathbf{R}_i^l \in \mathbb{R}^{3\times 3}$0 at point $\mathbf{R}_i^l \in \mathbb{R}^{3\times 3}$1, the method projects $\mathbf{R}_i^l \in \mathbb{R}^{3\times 3}$2 onto the corresponding FLAME triangle along the face normal and computes UV coordinates by barycentric interpolation from the triangle’s UVs. To avoid geometric projection and interpolation at every rendering event, the local mapping is condensed into a precomputable affine transform: $\mathbf{R}_i^l \in \mathbb{R}^{3\times 3}$3 Here, $\mathbf{R}_i^l \in \mathbb{R}^{3\times 3}$4 is the UV coordinate at the Gaussian center and $\mathbf{R}_i^l \in \mathbb{R}^{3\times 3}$5 is a transformation matrix precomputed at the Gaussian center.

The paper identifies an inconsistent-UV problem that arises when multiple unconstrained Gaussians intersect a ray and map to different UV locations; gradients then spread across multiple texture coordinates and the learned texture becomes blurred. To avoid this, surf-GS are strictly constrained to the mesh surface and their rotations are aligned with the triangle frame rather than independently optimized. Concretely, a surface Gaussian center is represented directly by barycentric coordinates on a triangle,

$\mathbf{R}_i^l \in \mathbb{R}^{3\times 3}$6

$\mathbf{R}_i^l \in \mathbb{R}^{3\times 3}$7

where $\mathbf{R}_i^l \in \mathbb{R}^{3\times 3}$8 are the triangle vertices and $\mathbf{R}_i^l \in \mathbb{R}^{3\times 3}$9 are learnable barycentric coordinates. This constraint is not a minor implementation detail; it is what makes sharp texture maps and reliable texture-space editing possible.

4. Volumetric branch and hybrid rendering

Surface Gaussians alone are not sufficient for the full head. Hair extends off the FLAME scalp surface and exhibits complex light transport, and lips are explicitly highlighted as difficult regions. SVG-Head therefore introduces vol-GS with a more flexible parameterization. These Gaussians still inherit the FLAME-bound local-to-global transport, so they remain animatable, but unlike surf-GS they store their own color directly and all their geometric parameters are optimized. For volumetric Gaussian $\boldsymbol{\mu}_i^l \in \mathbb{R}^3$0, the learnable parameters are local or global center $\boldsymbol{\mu}_i^l \in \mathbb{R}^3$1, opacity $\boldsymbol{\mu}_i^l \in \mathbb{R}^3$2, scale $\boldsymbol{\mu}_i^l \in \mathbb{R}^3$3, rotation $\boldsymbol{\mu}_i^l \in \mathbb{R}^3$4, and full SH coefficients $\boldsymbol{\mu}_i^l \in \mathbb{R}^3$5 (Sun et al., 13 Aug 2025).

Rendering follows standard 3D Gaussian Splatting, but the color function differs by Gaussian type. Vol-GS use directly stored SH-based colors. Surf-GS use UV-sampled texture values from $\boldsymbol{\mu}_i^l \in \mathbb{R}^3$6 together with the residual high-order SH component. The branches are then unified in the same alpha-compositing pipeline. The representation is therefore hybrid in both geometry and appearance: globally editable appearance is concentrated in texture space, while local residual appearance remains available in the Gaussian domain.

This design preserves compatibility with efficient splatting-based rasterization while changing the semantics of color evaluation. The paper refers to the resulting process as differential hybrid rendering: both branches are differentiable, jointly optimized, and rendered in one compositing system. A plausible implication is that SVG-Head shifts the main editable degrees of freedom from per-primitive color parameters to a low-dimensional, semantically organized UV domain, while reserving Gaussian flexibility for regions where UV-space appearance alone is insufficient.

5. Hierarchical optimization and loss design

SVG-Head is trained in two stages because jointly optimizing editable surf-GS and expressive vol-GS from scratch is under-constrained. In the first stage, only the surface branch and mesh-related variables are optimized. The learnable parameters include FLAME shape $\boldsymbol{\mu}_i^l \in \mathbb{R}^3$7, expression $\boldsymbol{\mu}_i^l \in \mathbb{R}^3$8, pose $\boldsymbol{\mu}_i^l \in \mathbb{R}^3$9; surf-GS barycentric coordinates $\mathbf{s}_i^l \in \mathbb{R}^3$0; surface opacity $\mathbf{s}_i^l \in \mathbb{R}^3$1; surface scales $\mathbf{s}_i^l \in \mathbb{R}^3$2; residual SH coefficients $\mathbf{s}_i^l \in \mathbb{R}^3$3; and the diffuse and dynamic textures $\mathbf{s}_i^l \in \mathbb{R}^3$4 and $\mathbf{s}_i^l \in \mathbb{R}^3$5. The image reconstruction loss is an $\mathbf{s}_i^l \in \mathbb{R}^3$6 plus D-SSIM combination,

$\mathbf{s}_i^l \in \mathbb{R}^3$7

and the method adds a diffuse-only rendering loss

$\mathbf{s}_i^l \in \mathbb{R}^3$8

together with a surface-scale regularizer

$\mathbf{s}_i^l \in \mathbb{R}^3$9

with $\mathbf{R}_j^t$0. The first-stage objective is

$\mathbf{R}_j^t$1

The diffuse-only term is important because it encourages a plausible explicit base texture suitable for editing, even when raw rendering metrics alone would not expose deficiencies in the learned texture image (Sun et al., 13 Aug 2025).

In the second stage, vol-GS are introduced for complementary modeling, but only a subset of the surface variables remains trainable: surf-GS opacity $\mathbf{R}_j^t$2 and dynamic texture $\mathbf{R}_j^t$3. Densification and pruning of surf-GS are halted. All vol-GS parameters are optimized. Additional regularizers are introduced. To keep volumetric Gaussians near their parent faces, the method uses

$\mathbf{R}_j^t$4

with $\mathbf{R}_j^t$5. To prevent surf-GS from becoming too transparent and thereby weakening the effect of editing, it imposes

$\mathbf{R}_j^t$6

where $\mathbf{R}_j^t$7 is the alpha map rendered using only surf-GS. The full second-stage objective is

$\mathbf{R}_j^t$8

The reported ablations support the staged schedule. Joint optimization from scratch performs noticeably worse. Freezing all surf-GS during stage two or unfreezing all of them also underperforms selective optimization of only $\mathbf{R}_j^t$9 and $\boldsymbol{\mu}_j^t$0. The paper interprets this as a balance issue: the surface representation must remain dominant enough to preserve explicit editability, but not so rigid that the volumetric branch cannot compensate for residual reconstruction errors.

6. Editing workflow, evaluation, and relation to adjacent methods

The editing interface follows directly from the representation. Because base appearance is encoded in explicit texture images, appearance editing amounts to editing the diffuse texture map $\boldsymbol{\mu}_j^t$1 directly in UV space. The paper describes localized changes such as adding makeup, recoloring parts of the face, or drawing fine-grained patterns, and emphasizes that no re-optimization is needed: once the texture is modified, the renderer immediately uses the updated texture values for surf-GS color lookup, and the edits remain consistent across views and expressions because the texture coordinates are tied to FLAME (Sun et al., 13 Aug 2025).

Experiments are reported on NeRSemble. The dataset provides multiview recordings from 16 viewpoints; images are downsampled to $\boldsymbol{\mu}_j^t$2, and the train/test protocol uses 9 of 10 expression sequences and 15 of 16 cameras for training, with the remaining sequence and camera reserved for evaluation. The baselines are PointAvatar, Gaussian Head Avatar (with super-resolution), GaussianAvatars, and MeGA. On novel-view synthesis, SVG-Head reports PSNR $\boldsymbol{\mu}_j^t$3, SSIM $\boldsymbol{\mu}_j^t$4, and LPIPS $\boldsymbol{\mu}_j^t$5. On novel-expression synthesis, it reports PSNR $\boldsymbol{\mu}_j^t$6, SSIM $\boldsymbol{\mu}_j^t$7, and LPIPS $\boldsymbol{\mu}_j^t$8. GaussianAvatars remains slightly better on the reported novel-view metrics at $\boldsymbol{\mu}_j^t$9, but it does not provide explicit real-time texture editing. MeGA is described as much worse in the reported setting, partly because poor shoulder modeling affects full-image metrics. The paper’s broader claim is therefore not that SVG-Head dominates all reconstruction metrics, but that it provides a distinct trade-off: competitive fidelity together with explicit texture images and real-time appearance editing.

Within the broader literature, SVG-Head sits between several adjacent design lines. MixedGaussianAvatar uses a surface-aligned field of 2D Gaussians attached to FLAME triangles and adds 3D Gaussian children only where appearance compensation is needed, making it a geometry-first surface-volumetric design rather than an explicitly texture-editable one (Chen et al., 2024). TexAvatars is another close conceptual neighbor: Gaussian attributes are predicted densely in UV space, but are lifted into 3D through mesh-aware Jacobians, emphasizing rigging stability and extrapolation under extreme deformations (Lee et al., 24 Dec 2025). MATCH likewise organizes Gaussian splat textures by a fixed surface parameterization and predicts them feed-forward in the fixed UV layout of a template mesh in $$\boldsymbol \mu_i = k_j \mathbf R^t_j \boldsymbol \mu^l_i+\boldsymbol \mu^t_j,$$0 seconds per frame, but it is not presented as a classic surface-volumetric hybrid (Prinzler et al., 16 Mar 2026). SurFhead moves in a different direction, replacing volumetric 3D Gaussians with 2D Gaussian surfels and affine deformation transfer to prioritize geometric reliability, normals, and depth (Lee et al., 2024).

This comparison clarifies the specific identity of SVG-Head. It is neither a purely surface-consistent surfel method nor a dense UV-registered Gaussian texture model, and it is not a free volumetric Gaussian cloud. Its distinctive contribution is the combination of FLAME-bound surface Gaussians, explicit texture images, and residual volumetric Gaussians in a representation designed not only for animatable rendering but also for direct, real-time appearance manipulation.