Research map / Point clouds / Shape and surface completion
Scaling Diffusion Mamba with Bidirectional SSMs for Efficient 3D Shape 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 DiM-3D, a diffusion model for 3D point cloud/voxel shape generation built on bidirectional Mamba SSM blocks instead of self-attention, reporting better ShapeNet metrics than DiT-3D with lower Gflops.
Problem
Diffusion transformers for 3D shape generation face quadratic attention cost, which becomes prohibitive at high voxel resolutions; the paper asks whether Mamba-style state space models can provide efficient, scalable 3D diffusion.
Approach
Replace self-attention in DiT-3D with bidirectional Mamba (SSM) blocks operating on patchified voxel tokens with a devoxelization output layer, giving linear complexity in sequence length for diffusion-based point cloud generation.
Key results
- On ShapeNet Chair, DiM-3D reaches 1-NNA CD 45.78 vs 49.11 for DiT-3D-XL (lower is better).
- At 256x256 voxel size, DiM-3D-XL/2 needs 294.58 Gflops vs 343.28 for DiT-3D-XL/2; at 2048x2048 DiT-3D runs out of memory while DiM-3D needs 18798.04 Gflops.
- Point cloud completion CD on Chair: 1.786 for DiM-3D vs 2.216 for DiT-3D.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace DiT-3D attention with the bidirectional Mamba-based DiM-3D architecture | DiT-3D · ShapeNet Chair, Airplane, and Car | 1-NNA CD: 49.11 -> 45.78; 62.35 -> 59.72; 48.24 -> 45.79 | Helped |
| replace DiT-3D attention with the bidirectional Mamba-based DiM-3D architecture | DiT-3D · ShapeNet Chair, Airplane, and Car | COV CD: 52.45 -> 57.89; 53.16 -> 60.67; 50.00 -> 59.63 | Helped |
| use DiM-3D for conditional point-cloud completion instead of DiT-3D | DiT-3D · ShapeNet Chair, Airplane, and Car completion | CD: 2.216 -> 1.786; 0.3521 -> 0.2516; 1.126 -> 0.832 | Helped |
| use DiM-3D for conditional point-cloud completion instead of DiT-3D | DiT-3D · ShapeNet Chair, Airplane, and Car completion | EMD: 2.385 -> 1.763; 0.9235 -> 0.7859; 1.513 -> 0.956 | Helped |
| use the DiM-3D Mamba architecture instead of DiT-3D | DiT-3D-XL/2 · 256x256, 512x512, and 1024x1024 voxel sizes | Gflops: 343.28 -> 294.58; 1371.08 -> 1175.7; 5482.28 -> 4700.17 | Helped |
| use DiM-3D-XL/2 instead of DiT-3D-XL/2 | DiT-3D-XL/2 · 2048x2048 voxel size | Gflops: OOM -> 18798.04 | Helped |
| increase model size from DiM-3D-S to DiM-3D-XL (ablation) | DiM-3D · ShapeNet Chair | 1-NNA CD: 57.27 -> 45.78 | Helped |
Surprising findings
- The authors report that direct extension of Mamba to point clouds does not work, requiring their voxel-based design.
Limitations
- Computational demands still significant for extremely high-resolution voxels and large datasets.
- Dependence on high-quality, diverse training data may limit generalization.
Experimental setup
Datasets: ShapeNet ShapeNet-55
Benchmarks: ShapeNet Chair ShapeNet Airplane ShapeNet Car ShapeNet-55 Mug and Bottle
Metrics: 1-NNA COV Chamfer Distance Earth Mover's Distance Gflops
Baselines: DiT-3D FastDiT-3D PVD LION MeshDiffusion Point-E PointFlow DPM
Builds on: DiT-3D Mamba DDPM PointFlow
Closest papers by meaning
- Direct3D-S2: Gigascale 3D Generation Made Easy with Spatial Sparse Attention (NeurIPS 2025 · similarity 0.88)
- LT3SD: Latent Trees for 3D Scene Diffusion (CVPR 2025 · similarity 0.88)
- DiMSUM: Diffusion Mamba - A Scalable and Unified Spatial-Frequency Method for Image Generation (NeurIPS 2024 · similarity 0.87)
- Skip Mamba Diffusion for Monocular 3D Semantic Scene Completion (AAAI 2025 · similarity 0.87)
- DiC: Rethinking Conv3x3 Designs in Diffusion Models (CVPR 2025 · similarity 0.87)
- Voxel Mamba: Group-Free State Space Models for Point Cloud based 3D Object Detection (NeurIPS 2024 · 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.