Research map / 3D vision / Gaussian splatting
Creating Your Editable 3D Photorealistic Avatar with Tetrahedron-constrained Gaussian Splatting
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.
Proposes TetGS, a hybrid tetrahedron-constrained Gaussian Splatting representation that enables localized, geometry-adaptive, photorealistic text- and image-guided editing of 360° full-body avatars from monocular videos.
Problem
Editing reconstructed 3D human avatars from monocular video with precise local region control, geometric adaptability, and photorealistic rendering; direct 3DGS editing under diffusion guidance yields unstable geometry and artifacts.
Approach
Embed Gaussian kernels inside tetrahedral grids (DMTet-style SDF on tetrahedron vertices, marching tetrahedra mesh) so tetrahedron updates guide Gaussian redistribution; decouple editing into localized spatial adaptation with partitioned tetrahedra and dual SDS losses, followed by coarse-to-fine texture generation using few-shot normal-based inpainting and attribute activation.
Key results
- Ours achieves FID 115.95 vs 194.54 (GaussianEditor), 201.82 (DGE), 258.27 (TIP-Editor) on rendered edited images.
- CLIP text-image similarity 26.28 for Ours vs 22.14 (GaussianEditor) and 23.05 (DGE); DINO 0.752 vs 0.726 for TIP-Editor.
- Localized spatial adaptation for 10000 iterations takes about 1.2 hours on a single NVIDIA A40.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TetGS-based decoupled editing instead of direct Gaussian editing with iN2N supervision (ablation) | 3D avatar editing · 3D avatar editing | FID: 215.56 -> 115.95 | Helped |
| Use TetGS-based decoupled editing instead of direct Gaussian editing with iN2N supervision (ablation) | 3D avatar editing · 3D avatar editing | CLIP: 22.64 -> 26.28 | Helped |
| Use TetGS-based decoupled editing instead of direct Gaussian editing with SDS supervision (ablation) | 3D avatar editing · 3D avatar editing | FID: 217.15 -> 115.95 | Helped |
| Use TetGS-based decoupled editing instead of direct Gaussian editing with SDS supervision (ablation) | 3D avatar editing · 3D avatar editing | CLIP: 23.96 -> 26.28 | Helped |
| Add attribute activation with augmented guidance (ablation) | TetGS editing pipeline · 3D avatar editing | FID: 120.33 -> 115.95 | Helped |
| Add attribute activation with augmented guidance (ablation) | TetGS editing pipeline · 3D avatar editing | CLIP: 26.03 -> 26.28 | Helped |
| Use TetGS instead of GaussianEditor | 3D avatar editing · 3D avatar editing (60 rendered images) | FID: 194.54 -> 115.95 | Helped |
| Use TetGS instead of GaussianEditor | 3D avatar editing · 3D avatar editing (60 rendered images) | CLIP: 22.14 -> 26.28 | Helped |
| Use TetGS instead of TIP-Editor | 3D avatar editing · 3D avatar editing (60 rendered images) | DINO: 0.726 -> 0.752 | Helped |
| Add tetrahedron partitioning (ablation) | Localized spatial adaptation · 3D avatar editing | rendering quality | Helped |
| Add local SDS loss (ablation) | Localized spatial adaptation · 3D avatar editing | detail quality | Helped |
| Add global SDS loss (ablation) | Localized spatial adaptation · 3D avatar editing | shape quality | Helped |
Limitations
- Evaluated on only 10 self-collected monocular videos
- Relies on multi-view masks from segmentation or human parsing models
- Few-shot inpainting may degrade novel-view quality, requiring attribute activation refinement
Experimental setup
Datasets: self-collected 360-degree full-body monocular video dataset (10 videos)
Benchmarks: self-collected 360-degree full-body avatar dataset
Metrics: FID CLIP score DINO similarity
Baselines: GaussianEditor DGE TIP-Editor
Builds on: 3D Gaussian Splatting DMTet NeuDA Instruct-NeRF2NeRF IDM-VTON
Closest papers by meaning
- 3DGS-Avatar: Animatable Avatars via Deformable 3D Gaussian Splatting (CVPR 2024 · similarity 0.90)
- SplattingAvatar: Realistic Real-Time Human Avatars with Mesh-Embedded Gaussian Splatting (CVPR 2024 · similarity 0.90)
- GASP: Gaussian Avatars with Synthetic Priors (CVPR 2025 · similarity 0.89)
- GaussianAvatars: Photorealistic Head Avatars with Rigged 3D Gaussians (CVPR 2024 · similarity 0.89)
- Human Gaussian Splatting: Real-time Rendering of Animatable Avatars (CVPR 2024 · similarity 0.88)
- GAvatar: Animatable 3D Gaussian Avatars with Implicit Mesh Learning (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.