Research map / 3D vision / Face and shape
Expressive Gaussian Human Avatars from Monocular RGB Video
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Builds EVA, a 3D Gaussian Splatting avatar driven by SMPL-X that learns expressive hand and face detail from monocular in-the-wild RGB video, using a fitting-based SMPL-X alignment module, context-aware adaptive density control, and a confidence-aware loss.
Problem
Learning expressive human avatars (fine hand and facial detail) from monocular in-the-wild RGB video, where off-the-shelf SMPL-X estimates are misaligned with frames and body-focused Gaussian avatar methods under-model small body parts.
Approach
Combine 3D Gaussians anchored to SMPL-X vertices with (i) a multi-stage keypoint/prior-based SMPL-X alignment step, (ii) densification thresholds that depend on body-part attribute and gradient history, and (iii) a per-pixel confidence weighting of the RGB loss predicted from rendered image and depth.
Key results
- On XHumans, EVA reaches 29.67 full-body PSNR and 33.05 full-body LPIPS, versus 29.16 / 41.16 for GauHuman + SMPL-X.
- On UPB, EVA improves hand LPIPS to 45.90 from 61.44 for the best baseline GART + SMPL-X (over 25% relative gain).
- On UPB, removing the SMPL-X alignment module raises full-body LPIPS from 65.37 to 73.82.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use context-aware adaptive density control instead of original density control (ablation) | EVA · XHumans full | LPIPS: 35.14 -> 33.05 | Helped |
| Add confidence weighting to the RGB loss (ablation) | EVA · XHumans full | LPIPS: 34.08 -> 33.05 | Helped |
| Add the SMPL-X alignment module instead of using the off-the-shelf estimate (ablation) | EVA · UPB full | LPIPS: 73.82 -> 65.37 | Helped |
Limitations
- Non-rigid elements such as cloth and hair are not modeled.
- Occlusion and few-shot settings are left for future work.
Experimental setup
Datasets: XHumans UPB
Benchmarks: XHumans UPB
Metrics: PSNR SSIM LPIPS
Baselines: 3DGS + SMPL-X GART + SMPL-X GauHuman + SMPL-X
Builds on: 3D Gaussian Splatting SMPL-X GART GauHuman
Closest papers by meaning
- Real-time High-fidelity Gaussian Human Avatars with Position-based Interpolation of Spatially Distributed MLPs (CVPR 2025 · similarity 0.89)
- Vid2Avatar-Pro: Authentic Avatar from Videos in the Wild via Universal Prior (CVPR 2025 · similarity 0.89)
- Gaussian Eigen Models for Human Heads (CVPR 2025 · similarity 0.89)
- Human Gaussian Splatting: Real-time Rendering of Animatable Avatars (CVPR 2024 · similarity 0.89)
- GASP: Gaussian Avatars with Synthetic Priors (CVPR 2025 · similarity 0.88)
- RAM-Avatar: Real-time Photo-Realistic Avatar from Monocular Videos with Full-body Control (CVPR 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.