GTAvatar
Bridging Gaussian Splatting and Texture Mapping for Relightable and Editable Gaussian Avatars
Eurographics 2026
Kelian Baert Mae Younes Francois Bourel Marc Christie Adnane Boukhayma
Paper Code (coming soon) Results BibTeX
GTAvatar broadens applications of monocular Gaussian Splatting head avatars beyond reenactment and relighting, enabling interactive editing through textures. From a single video, we reconstruct a state-of-the-art 3D avatar, integrating material textures while preserving the training efficiency, rendering speed and visual fidelity of Gaussian Splatting.
Overview
We leverage the explicit surface representation of 2D Gaussians to map the ray-splat intersection to UV space:
This mapping is computed efficiently as detailed in our paper. From this UV coordinate, we retrieve texel values for albedo, roughness, specular reflectance and normal via bilinear texture filtering. These values replace the Spherical Harmonics used in standard Gaussian Splatting. Finally, the final image is rendered using the Cook-Torrance BRDF with the rasterized G-buffers. The Gaussian attributes and texture maps are jointly optimized end-to-end during training, with photometric, geometric and regularization losses.
Gaussian-to-Mesh binding
Each Gaussian is embedded onto a triangle of the FLAME 3D Morphable Model with a fixed triangle ID, and parameterized by barycentric coordinates within the triangle, an offset along the normal, a relative rotation, two scales and an opacity. Given pose and expression parameters, Gaussian attributes are updated as follows:
- Deform mesh,
- Compute 3D position using the triangle's vertices, the barycentric coordinates and the offset along the normal,
- Multiply the splat's relative rotation by the triangle's orientation matrix (TBN).
- Compute uv0, the UV coordinates at the center of the splat.
- Compute uvJ, the Jacobian matrix used in the mapping (see Section 3.4 of the paper for details).
During initialization, Gaussians are created at the center of each triangle. Typical densification and pruning from 2DGS are then used during optimization to refine the distribution of Gaussians over the surface.
Qualitative Results
Texture editing
The semantic texture space enables intuitive editing of the avatars' appearance with widely available tools. This section illustrates applications of our method through simple modifications.
Albedo editing
Here, the albedo texture is replaced or edited manually in image editing software. The modified texture integrates with the original reconstruction seemlessly, preserving high-frequency details. Note that those edits are fully compatible with animation and relighting.
Material editing
In the following examples, we partially or completely replace the material maps, resulting in substantial changes to the appearance. As shown, normal mapping can be used to add intricate details to the geometry. Our pipeline is compatible with off-the-shelf material definitions.
Reconstruction
The following avatars are rendered with head poses and expressions tracked from video frames that were not seen during training (self-reenactment).
Relighting
In addition to new poses and expressions, we modify the illumination of the scene using environment maps. Our approach achieves results that are comparable to state-of-the-art relighting methods.
For more examples along with comparisons, quantitative results and ablation studies, please refer to the paper.
BibTeX
If you find this work useful for your research, please cite:
(coming soon)
@article{baert2026gta,
author = {},
title = {},
journal = {},
number = {},
volume = {},
month = {},
year = {},
url = {}
}