Research map / 3D vision / Gaussian splatting
Tetrahedron Splatting for 3D Generation
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 TeT-Splatting, a 3D representation that splats alpha-blended tetrahedra from a deformable tetrahedral grid with SDF-based opacity, enabling real-time rendering, precise Marching Tetrahedra mesh extraction, and faster convergence for 3D generation.
Problem
3D generation with 2D diffusion priors needs a representation that converges quickly, renders in real time, and yields accurate meshes; NeRF is costly and inaccurate for meshes, DMTet converges poorly on large topology changes, and 3DGS yields poor meshes with holes.
Approach
Treat tetrahedra of a deformable tetrahedral grid as splatting primitives: project them to 2D, compute opacity from SDF values at ray-tetrahedron intersections (NeuS-style sigmoid), alpha-blend in a tile-based CUDA rasterizer, and regularize the SDF with eikonal and normal-consistency losses so the result can be meshed with Marching Tetrahedra.
Key results
- Geometry CLIP score 23.16 and appearance CLIP score 29.42 versus RichDreamer's 23.01 and 29.22 under the authors' re-run on the 183-object prompt list.
- Geometry stage with rich priors takes about 40 min versus 70 min for RichDreamer.
- Rendering runs in real time but at lower FPS than DMTet (Fantasia3D) and 3DGS (DreamGaussian).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TeT-Splatting instead of RichDreamer | RichDreamer text-to-3D pipeline with rich diffusion priors · RichDreamer prompts (183 objects) | Geometry CLIP score: 23.0143 -> 23.1641 | Helped |
| Use TeT-Splatting instead of RichDreamer | RichDreamer text-to-3D pipeline with rich diffusion priors · RichDreamer prompts (183 objects) | Appearance CLIP score: 29.2198 -> 29.4197 | Helped |
| Use TeT-Splatting instead of RichDreamer | RichDreamer text-to-3D pipeline with rich diffusion priors · Text-to-3D geometry optimization | optimization time: 70 min -> 40 min | Helped |
| Remove eikonal loss (ablation) | TeT-Splatting · Text-to-3D with rich diffusion priors | shape quality | Hurt |
| Add eikonal loss (ablation) | TeT-Splatting · Text-to-3D with rich diffusion priors | shape quality | Helped |
| Remove normal consistency loss (ablation) | TeT-Splatting · Text-to-3D with rich diffusion priors | model compactness | Hurt |
| Add normal consistency loss (ablation) | TeT-Splatting · Text-to-3D with rich diffusion priors | surface fragmentation | Helped |
| Increase tetrahedral grid resolution from 128 to 256 (ablation) | TeT-Splatting · Text-to-3D geometry generation | geometry detail: 128 -> 256 | Helped |
Surprising findings
- Under the authors' own prompt list, RichDreamer's reported geometry CLIP score (25.88 in their paper) is higher than TeT-Splatting's, so the geometry CLIP advantage only holds when comparing against their re-run.
Limitations
- Tetrahedra as primitives limit high-frequency detail to the tetrahedral grid resolution, so texture is handled by a separate mesh stage.
- Rasterizer is slower than 3DGS; only pre-filtering is used without a densification process.
- Lower FPS than DMTet-based and 3DGS-based baselines.
Experimental setup
Datasets: prompts_dmtet.txt (183 objects, from RichDreamer GitHub)
Benchmarks: CLIP score (geometry and appearance) on 183-object prompt list qualitative text-to-3D and image-to-3D comparisons FPS and training time comparisons
Metrics: Geometry CLIP score Appearance CLIP score rendering FPS training time
Baselines: Magic3D Fantasia3D DreamGaussian Instant-NGP ProlificDreamer RichDreamer MVDream
Builds on: DMTet 3D Gaussian Splatting NeuS Voxurf Fantasia3D
Closest papers by meaning
- TetSphere Splatting: Representing High-Quality Geometry with Lagrangian Volumetric Meshes (ICLR 2025 · similarity 0.92)
- DirectTriGS: Triplane-based Gaussian Splatting Field Representation for 3D Generation (CVPR 2025 · similarity 0.89)
- SuGaR: Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering (CVPR 2024 · similarity 0.89)
- DiffGS: Functional Gaussian Splatting Diffusion (NeurIPS 2024 · similarity 0.88)
- Creating Your Editable 3D Photorealistic Avatar with Tetrahedron-constrained Gaussian Splatting (CVPR 2025 · similarity 0.87)
- 3D Convex Splatting: Radiance Field Rendering with 3D Smooth Convexes (CVPR 2025 · similarity 0.87)
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.