Research map / 3D vision / 3D generation
Ultra3D: Efficient and High-Fidelity 3D Generation with Part Attention
Yiwen Chen, Zhihao Li, Yihao Luo, Yikai Wang, Zhang Hu, Le Li, Qin Li, Chi Zhang, Guosheng Lin
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace full attention in the second-stage DiT with Part Attention (ablation) | Ultra3D · User study (b) full vs part attention | selection rate (%): 12.4% vs 8.9%, 78.7% no preference | No clear effect |
| use Part Attention instead of fixed 3D window attention (ablation) | Ultra3D · User study (c) 3D window vs part attention | selection rate (%): 2.1% -> 63.7% | Improved |
| use Part Self Attention instead of full attention (ablation) | Ultra3D DiT · Efficiency vs full attention | speedup: 6.7x | Improved |
| use Part Cross Attention instead of full cross attention (ablation) | Ultra3D DiT · Efficiency vs full attention | speedup: 4.1x | Improved |
| replace most DiT layers with Part Attention (ablation) | Ultra3D DiT · Efficiency vs full attention | training and inference speedup: 3.1x training, 3.3x inference | Improved |
| use Ultra3D with Part Attention instead of full attention pipeline (ablation) | Ultra3D · Per-sample generation time | minutes per mesh: over 15 minutes -> 4 minutes | Improved |
| compare Ultra3D against Direct3D-S2 and Commercial Model A | Ultra3D · User study (a) vs other methods | selection rate (%): 7.2% and 24.3% -> 68.5% | Improved |
| vary number of part groups at inference away from the 8 used in training (ablation) | Ultra3D · Qualitative generation quality | visual quality | No clear effect |
| increase sparse voxel resolution used in attention and drop downsampling (ablation) | Ultra3D · Qualitative generation quality | visual quality | Improved |
Closest papers by meaning
- Direct3D-S2: Gigascale 3D Generation Made Easy with Spatial Sparse Attention (NeurIPS 2025 · similarity 0.90)
- PartCrafter: Structured 3D Mesh Generation via Compositional Latent Diffusion Transformers (NeurIPS 2025 · similarity 0.90)
- PartGen: Part-level 3D Generation and Reconstruction with Multi-view Diffusion Models (CVPR 2025 · similarity 0.88)
- Sculpt3D: Multi-View Consistent Text-to-3D Generation with Sparse 3D Prior (CVPR 2024 · similarity 0.87)
- MoCA: Mixture-of-Components Attention for Scalable Compositional 3D Generation (NeurIPS 2026 · similarity 0.87)
- SILSA: Sliding-Window Slice Latents for Topology-Preserving High-Resolution 3D Generation (NeurIPS 2026 · 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.
