Research map / Point clouds / Shape and surface completion
Fast Training of Diffusion Transformer with Extreme Masking for 3D Point Clouds Generation
Shentong Mo, Enze Xie, Yue Wu, Junsong Chen, Matthias Niessner, Zhenguo Li
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 |
|---|---|---|---|
| add 3D voxel masking (ablation) | DiT-3D baseline · ShapeNet Chair, voxel 32 | training cost (hours): 91 -> 11 | Improved |
| add 3D voxel masking (ablation) | DiT-3D baseline · ShapeNet Chair, voxel 32 | 1-NNA@CD: 51.99 -> 50.09 | Improved |
| add 3D window attention decoder (ablation) | FastDiT-3D with voxel masking · ShapeNet Chair, voxel 32 | training cost (hours): 11 -> 8 | Improved |
| add window attention decoder on quality metrics (ablation) | FastDiT-3D with voxel masking · ShapeNet Chair, voxel 32 | 1-NNA@CD: 50.09 -> 50.35 | No clear effect |
| increase random masking ratio to 99% (ablation) | FastDiT-3D · ShapeNet Chair, voxel 32 | 1-NNA@CD: 51.53 -> 82.35 (95% to 99%) | Hurt |
| use foreground-background aware masking instead of random masking at 95% (ablation) | FastDiT-3D · ShapeNet Chair, voxel 32 | 1-NNA@CD: 51.53 -> 50.22 | Improved |
| increase non-occupied masking ratio from 99% to 100% (ablation) | voxel-aware masking · ShapeNet Chair, voxel 32 | 1-NNA@CD: 50.09 -> 52.87 | Hurt |
| add mixture-of-experts FFN (k=2) in encoder (ablation) | FastDiT-3D · ShapeNet Chair, voxel 32 | 1-NNA@CD: 52.16 -> 51.72 | Improved |
| initialize from ImageNet pre-trained MaskDiT weights (ablation) | FastDiT-3D · ShapeNet Chair, voxel 32 | training cost (hours): 8 -> 7 hours; 1-NNA@CD 50.35 -> 50.39 | No clear effect |
| use FastDiT-3D instead of DiT-3D at voxel size 128 | 3D point cloud generation · ShapeNet Chair, voxel 128 | training cost (A100 hours): 1668 -> 108 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- DiT-3D — DiT-3D: Exploring Plain Diffusion Transformers for 3D Shape Generation (NeurIPS 2023) · better on 8 of 8 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- TIGER: Time-Varying Denoising Model for 3D Point Cloud Generation with Diffusion Process (CVPR 2024 · similarity 0.89)
- MeshFormer : High-Quality Mesh Generation with 3D-Guided Reconstruction Model (NeurIPS 2024 · similarity 0.89)
- Scaling Diffusion Mamba with Bidirectional SSMs for Efficient 3D Shape Generation (AAAI 2025 · similarity 0.89)
- Denoising Diffusion via Image-Based Rendering (ICLR 2024 · similarity 0.89)
- SeaLion: Semantic Part-Aware Latent Point Diffusion Models for 3D Generation (CVPR 2025 · similarity 0.88)
- PointInfinity: Resolution-Invariant Point Diffusion Models (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.
