Research map / 3D vision / Face and shape
GALA: Geometry-Aware Local Adaptive Grids for Detailed 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 GALA, a sparse forest-of-octrees 3D shape representation with oriented, anisotropically scaled local SDF grids, plus a cascaded diffusion pipeline for generating shapes from it.
Problem
Efficient, accurate 3D shape representations that capture fine geometric detail with few parameters and that are suitable for diffusion-based 3D generation.
Approach
Place a forest of octrees only on the object surface, and in each non-empty leaf fit a local SDF grid whose orientation comes from PCA of local normals and whose axes are anisotropically rescaled using a histogram of sampled points; the flattened grid configurations are then generated cascade-style with transformers and diffusion.
Key results
- Reconstruction on 50 ShapeNet objects: CD 1.18e-3 and LFD 1109.36 with 277K parameters, versus 2.61e-3 CD and 1.05M parameters for MeshDiffusion.
- Fitting completes in under 10 seconds per object on a single A100 (extraction 2.56±1.30s on ShapeNet Airplane), with peak GPU memory under 800MB.
- Airplane generation MMD-CD 2.48 (x10^-3) and 1-NNA-CD 59.90%, outperforming Mosaic-SDF (3.54 and 62.62%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use local adaptive grids instead of uniform axis-aligned grid samplings (ablation) | GALA · ShapeNet reconstruction (5 classes) | CD (×10^-3): 1.52 -> 1.18 | Helped |
| add bounded-normal orientation to vanilla grid extraction (ablation) | GALA · ShapeNet reconstruction test | CD (×10^-3): 1.55 -> 1.33 | Helped |
| add histogram rescaling to bounded-normal orientation (ablation) | GALA · ShapeNet reconstruction test | CD (×10^-3): 1.33 -> 1.18 | Helped |
| add bounded-normal orientation to vanilla grid extraction (ablation) | GALA · ShapeNet reconstruction test | HD (×10^-2): 1.46 -> 1.19 | Helped |
| add histogram rescaling to bounded-normal orientation (ablation) | GALA · ShapeNet reconstruction test | HD (×10^-2): 1.19 -> 1.06 | Helped |
| quantize the local adaptive grid quantities during fitting (ablation) | GALA with normals and histogram rescaling · ShapeNet reconstruction test | CD (×10^-3): 1.08 -> 1.18 | Hurt |
| vary the subdivision overlap ratio and use α=0.2 (ablation) | GALA · ShapeNet reconstruction test | Chamfer Distance | Helped |
| generate with GALA instead of the strongest reported baseline | ShapeNet airplane generation · ShapeNet airplane | MMD CD: 3.36 -> 2.48 | Helped |
| generate with GALA instead of Mosaic-SDF | ShapeNet chair generation · ShapeNet chair | COV EMD (%): 55.03 -> 53.10 | Hurt |
| generate with GALA instead of Mosaic-SDF | ShapeNet chair generation · ShapeNet chair | MMD CD: 15.47 -> 13.28 | Helped |
Surprising findings
- Adaptive grid orientation and anisotropic scaling allow 8-bit quantization with little loss (CD 1.18e-3 quantized vs 1.08e-3 unquantized).
Limitations
- Requires watertight input meshes.
- Generated shapes can have unconnected geometry when root voxel generation is imperfect.
- Mosaic-SDF comparison relies on results supplied by its authors, and the self-implemented baseline is approximate.
Experimental setup
Datasets: ShapeNet
Benchmarks: ShapeNet
Metrics: Chamfer Distance Light Field Distance Maximum Mean Discrepancy Coverage 1-nearest neighbor accuracy Earth Mover Distance fitting time parameter count
Baselines: 3DShape2VecSet Neural Wavelet MeshDiffusion Mosaic-SDF 3DILG DMTet
Builds on: Mosaic-SDF OBBTree DiT RePaint MeshGPT
Closest papers by meaning
- CraftsMan3D: High-fidelity Mesh Generation with 3D Native Diffusion and Interactive Geometry Refiner (CVPR 2025 · similarity 0.88)
- ARMesh: Autoregressive Mesh Generation via Next-Level-of-Detail Prediction (NeurIPS 2025 · similarity 0.87)
- 3DTopia-XL: Scaling High-quality 3D Asset Generation via Primitive Diffusion (CVPR 2025 · similarity 0.87)
- SILSA: Sliding-Window Slice Latents for Topology-Preserving High-Resolution 3D Generation (NeurIPS 2026 · similarity 0.87)
- Direct3D-S2: Gigascale 3D Generation Made Easy with Spatial Sparse Attention (NeurIPS 2025 · similarity 0.87)
- MeshGen: Generating PBR Textured Mesh with Render-Enhanced Auto-Encoder and Generative Data Augmentation (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.