Expressive Whole-Body 3D Gaussian Avatar

• Most casual-video avatars support body motion but not facial expressions or hands, while prior whole-body work required 3D observations and diverse registered expressions and poses.The cited prior work is.Source: §1 · p. 2
• ExAvatar combines SMPL-X and 3D Gaussian Splatting, treating each Gaussian as a surface vertex with triangle connectivity following SMPL-X topology.Source: §1 · p. 2
• The shared topology makes ExAvatar compatible with SMPL-X facial-expression codes, so novel expressions are not strictly limited by the number of training frames.The paper contrasts this with.Source: §1 · p. 3
• Connectivity-based regularizers encourage local similarity between Gaussians and significantly reduce artifacts in novel facial expressions and poses.Limited pose diversity and unobserved parts can leave geometry ambiguous.Source: §1 · p. 3
• ExAvatar substantially outperforms all previous 3D human avatars across various benchmarks.Source: §1 · p. 3

3D human avatars

• Avatar work ranges from capture-studio methods using accurate 3D or multiview supervision to monocular-video methods that avoid requiring those observations.Recent 3DGS-based avatars use approaches including triplanes and posed-SMPL positional maps.Source: §2 · pp. 3–4
• Most reviewed avatars support body motion but not facial expressions and hand motions; X-Avatar supports whole-body animation but requires diverse expressions and accurate 3D observations.Liu et al. [30] present another whole-body avatar, but it does not animate novel facial expressions.Source: §2 · p. 4

Whole-body 3D human modeling and perception

• SMPL-X is the most widely used among parametric whole-body models, which represent minimally clothed humans using pose, facial-expression, and shape parameters.These models describe human geometry; whole-body pose-estimation methods also build on them.Source: §2 · p. 4

3D human avatars

• The cited offset comparison identifies joint and face offsets as necessary for accurate hand and face registration and whole-body co-registration.Source: Figure 2 · p. 5

3.1 Accurate co-registration of SMPL-X

• Shared joint and face offsets refine SMPL-X registration, with joint offsets helping fit hands and face offsets fitted against FLAME face vertices.The joint offsets affect both skeleton and surface; face offsets are per-vertex adjustments in the T-pose face region.Source: §3.1 · pp. 5–6

3.2 Architecture

• A canonical mesh with 167K vertices and 335K faces anchors a triplane-based representation that regresses Gaussian assets for each vertex.Pose-conditioned MLPs predict geometry and scale offsets plus RGB offsets, using vertex normals for the color offsets.Source: §3.2 · pp. 6–7

3.3 Animation and rendering

• Animation adds SMPL-X facial-expression offsets to face vertices and applies pose-dependent offsets before linear blend skinning and 3DGS rendering.For body vertices, skinning weights come from nearest downsampled canonical-mesh vertices; hand and face vertices use their original weights.Source: §3.3 · p. 8

3.4 Loss functions

• Training jointly optimizes avatar parameters and frame-specific pose, expression, and translation, while a mesh-rendered face loss enforces face geometry-texture consistency.The face loss adjusts face Gaussian positions against a captured image using fixed UV texture from registered face-only models.Source: §3.4 · pp. 9–10
• Connectivity-based Laplacian regularization reduces artifacts in novel expressions and poses by encouraging local similarity among Gaussian points.It is reported as more effective than the L2 distance-to-template regularizer and is also applied to Gaussian scales and RGB values.Source: Figure 7 · p. 10, §3.4 · p. 11

4.1 Datasets

• NeuMan uses short in-the-wild monocular videos, while X-Humans provides studio scans and RGBD videos with more diverse expressions and hand poses.ExAvatar is created from monocular RGB video for the X-Humans comparison, against prior work using RGBD videos.Source: §4.1 · p. 11

4.2 Comparison to state-of-the-art methods

• ExAvatar achieves the best results on NeuMan with and without rendered backgrounds, and outperforms the prior whole-body avatar on X-Humans without depth maps.The X-Humans comparison uses a prior method trained with additional depth maps.Source: §4.2 · pp. 11–12, Table 1 · p. 12, Table 2 · p. 12
• ExAvatar produces sharper shirt prints and better faces and hands in novel views and poses than prior avatars in the qualitative comparison.The authors attribute the sharper face and hand textures to whole-body modeling and contrast them with blurred textures in prior avatars lacking face and hand controllability.Source: §4.2 · p. 12, Figure 8 · p. 13

4.3 Ablation study

• Ablations show that Laplacian regularization improves performance and stability, while face loss benefits evaluation on cropped images where the face is visible.Face visibility is determined by rasterizing SMPL-X meshes and checking the number of rasterized face-region triangles.Source: §4.3 · pp. 12–13, Table 4 · p. 13, Table 5 · p. 13

5 Conclusion

• The hybrid mesh and 3D Gaussian representation supports animation through the SMPL-X facial-expression space and reduces artifacts in novel expressions and poses.The approach addresses limited expression and pose diversity and the lack of 3D observations in short monocular-video training.Source: §5 · p. 13
• Unobserved mouth interiors and palms receive hallucinated geometry and textures, while dynamic clothes remain difficult to model.Modeling clothing material with motion information such as velocity and acceleration is outside the paper’s scope.Source: §5 · p. 14

Supplementary Material for “Expressive Whole-Body 3D Gaussian Avatar"

• The supplement covers SMPL-X co-registration, MLP architecture, regularizers, runtime, geometry, generative-AI comparisons, implementation, and failure cases.The supplement describes these as details and experiments omitted from the main text due to page limits.Source: Supplementary Material for “Expressive… · p. 15

A Details of co-registration of SMPL-X

• FLAME parameters from DECA are fitted to face keypoints using keypoint and initialization losses, aligning the mesh to the image while limiting drift from the regressed parameters.The initialization loss is weighted by 0.1.Source: A Details of co-registration of SMPL-X · p. 15
• SMPL-X registration fits Hybrid-X parameters to whole-body keypoints, substitutes DECA’s expression code, and uses joint and face offsets to personalize the model.The joint offset affects both skeleton and surface through forward kinematics and linear blend skinning.Source: A Details of co-registration of SMPL-X · p. 15
• The SMPL-X objective combines keypoint, initialization, face-alignment, and regularization losses; face alignment is optimized only when the face is visible.Face terms align SMPL-X face vertices, Laplacians, and edge lengths with FLAME; regularizers constrain shape and offsets, including their symmetry.Source: A Details of co-registration of SMPL-X · pp. 15–16

B Architecture of MLPs

• The supplementary section on MLP architecture refers readers to Fig. 4 in the main manuscript for its brief description.This passage does not specify the MLP architecture itself.Source: B Architecture of MLPs · p. 16

B.1 MLPs without pose conditioning

• Without pose conditioning, one MLP maps triplane features to Gaussian geometry and another maps them to RGB colors.Both use four fully connected layers with hidden size 128, group normalization, and ReLU; the color MLP differs in output dimension.Source: §B.1 · p. 16
• 47 FPS is reported for ExAvatar without pose conditioning, compared with 26 FPS for ExAvatar with pose conditioning.The comparison covers NeuMan (0.02), Vid2Avatar (0.04), and X-Avatar (0.12) as well.Source: §B.1 · p. 17

B.2 MLPs with pose conditioning

• Table B compares ExAvatar and X-Avatar on a subset of the X-Humans test set for 3D geometry.The supplied passage identifies the comparison scope but does not report its values.Source: §B.1 · p. 17
• With pose conditioning, geometry offsets use triplane features and SMPL-X pose, while color offsets also use each Gaussian's normal.The pose input excludes the root pose; the MLP architectures otherwise match those described in Sec. B.1.Source: §B.2 · p. 17

C Regularizers

• The regularizers penalize Gaussian offsets from the canonical mesh, large Gaussian scales, and deviations of hand Gaussian colors from their average.The section also uses the same Ljo as Sec. A.Source: C Regularizers · p. 17

D Running time comparison

• ExAvatar achieves real-time animation and rendering, while the cited existing works fail to do so; removing pose conditioning makes it faster.Timing covers human animation and rendering only, excluding background rendering, at 1024 × 1024 on one NVIDIA V100 GPU.Source: D Running time comparison · p. 17
• The appendix includes qualitative comparisons of rendered normals against X-Avatar and human animation against MagicAnimate.The captions identify the comparisons but do not state their outcomes.Source: D Running time comparison · p. 18

E Geometry comparison

• ExAvatar produces better reported 3D geometry than X-Avatar, especially around the face and neck.The comparison renders masks, depth maps, and normal maps from multiple viewpoints against ground truth; the authors leave actual geometry extraction for future work.Source: E Geometry comparison · p. 18
• The reported geometry comparison is for reference because screen-space depth and normal maps may be incorrect for elongated Gaussians.The described methods take depth or normals from Gaussian centers and alpha-blend them.Source: E Geometry comparison · p. 18
• A wrong facial-expression code from the off-the-shelf regressor is shown as an ExAvatar failure case.The figure caption attributes the case to the regressed expression code ψ.Source: Figure C · p. 19
• ExAvatar may lack plausibility for unobserved parts such as the inside of the mouth and cloth dynamics, while retaining more authenticity than the cited generative result.The authors identify combining generative plausibility with their approach's authenticity as future work.Source: F Comparison to generative AIs · p. 19

G Implementation details

• Training uses Adam with an initial learning rate of 10^-3, reduced by a factor of 10 at 75% and 90% of total iterations.PyTorch is used for implementation.Source: G Implementation details · p. 19
• Training runs on a single NVIDIA V100 GPU and takes 1.5 to 5 hours depending on video duration.The authors state that other implementation details will be available in their code.Source: G Implementation details · p. 19
• The authors state that additional implementation details will be available in their code.Source: G Implementation details · p. 19

H Failure cases

• An incorrect facial-expression code from the off-the-shelf regressor makes both the rendered SMPL-X mesh and ExAvatar show the wrong expression.The authors suggest investigating advanced facial models and high-fidelity regressors to address this failure case.Source: H Failure cases · p. 20
• A subject who always smiles during the short NeuMan capture has smiling facial appearance baked into the avatar, which renders slightly smiling when the subject frowns.When the face occupies only a few pixels, noisy 2D keypoints can misalign SMPL-X/FLAME lip geometry during co-registration, and that error propagates into ExAvatar learning.Source: H Failure cases · p. 20
• The authors identify generative priors for canonical facial appearance as one possible way to address the baked-expression failure case.They also note the challenge is especially acute when the face occupies only a few pixels in the video.Source: H Failure cases · p. 20