Research map / 3D vision / Gaussian splatting
Hierarchical Gaussian Mixture Model Splatting for Efficient and Part Controllable 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 HGMM splatting with a tree-topology Diffusion Mamba and cascaded latent-then-Gaussian diffusion for fast, part-controllable 3D Gaussian generation from text or images.
Problem
Existing Gaussian-splatting 3D generators need complex preprocessing for disordered Gaussian primitives, lack part-level representation, and offer limited controllability over generation.
Approach
Fit a fixed number of Gaussians with a hierarchical GMM tree (explicit Gaussians plus implicit latents per level), model the disordered Gaussians with a Mamba diffusion network that scans them along a minimum spanning tree built by BFS/Boruvka, and generate the hierarchical latents first in a cascade before splatting parameters.
Key results
- Unconditional generation on ShapeNet Car achieves FID 11.03 vs. 13.01 for GaussianCube; OmniObject3D class-conditioned FID 10.57 vs. 11.62.
- Text-to-3D CLIP score 30.61 at about 5 s inference on a Tesla V100, versus 30.52 for Shap-E and 30.06 for LGM.
- Image-to-3D on 3DBiCar reaches PSNR 27.63 vs. 24.78 for GaussianCube; user study prefers the cascaded model for text alignment 87.27% of the time.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase the primitive count from 16,384 to 36,864 (ablation) | HGMM Splatting · ShapeNet Chair fitting | PSNR: 32.74 -> 34.99 | Helped |
| Increase the primitive count from 36,864 to 65,536 (ablation) | HGMM Splatting · ShapeNet Chair fitting | PSNR: 34.99 -> 35.01 | Helped |
| Replace tree scanning with selective scanning (ablation) | Diffusion Mamba · ShapeNet Car generation | FID-50K: 12.53 -> 43.56 | Hurt |
| Replace tree scanning with PCM scanning (ablation) | Diffusion Mamba · ShapeNet Car generation | FID-50K: 12.53 -> 52.69 | Hurt |
| Use BFS initialization for tree scanning (ablation) | Diffusion Mamba · ShapeNet Car generation | PSNR | Helped |
| Add cascaded diffusion for hierarchical latent generation (ablation) | Our Method · Text-prompt alignment | user preference alignment: 12.73% -> 87.27% | Helped |
Limitations
- Relies on ShapeNet, OmniObject3D, and a filtered Objaverse subset; no large-scale pretrained models are used.
- Controllability evaluated mainly through a user study and qualitative examples rather than quantitative part-level metrics.
Experimental setup
Datasets: ShapeNet Car ShapeNet Chair OmniObject3D Objaverse LVIS subset 3DBiCar
Benchmarks: ShapeNet Car ShapeNet Chair OmniObject3D Objaverse 3DBiCar
Metrics: FID KID PSNR SSIM LPIPS CLIP score MMD COV FPD IoU
Baselines: GaussianCube DiffTF EG3D GET3D DreamGaussian Shap-E LGM
Builds on: GaussianCube Mamba GrootVL 3D Gaussian Splatting LGM
Closest papers by meaning
- Text-to-3D using Gaussian Splatting (CVPR 2024 · similarity 0.91)
- Lightweight Predictive 3D Gaussian Splats (ICLR 2025 · similarity 0.90)
- GaussianDreamer: Fast Generation from Text to 3D Gaussians by Bridging 2D and 3D Diffusion Models (CVPR 2024 · similarity 0.90)
- 3D Gaussian Splatting as Markov Chain Monte Carlo (NeurIPS 2024 · similarity 0.90)
- DiffSplat: Repurposing Image Diffusion Models for Scalable Gaussian Splat Generation (ICLR 2025 · similarity 0.90)
- DirectTriGS: Triplane-based Gaussian Splatting Field Representation for 3D Generation (CVPR 2025 · similarity 0.90)
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.