Research map / 3D vision / 3D generation
Position: You Can't Manufacture a NeRF
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.
The position paper argues that generative 3D models (NeRFs, meshes, primitive-to-CAD) fail manufacturability requirements, and releases a relabeled F360 segmentation dataset plus Thang3D real-part set with a ParSeNet decomposition baseline.
Problem
Generative 3D models produce outputs (meshes, NeRFs, point-based CAD reconstructions) that are not watertight, manifold, or precise enough for real-world CNC and mass manufacturing, and do not generalize beyond their training data.
Approach
Argues that manufacturability requires fine-grained, classless geometric decomposition into BREP-level faces rather than class-based primitive featurization; shows existing models fail on realistic data and provides a relabeled F360 dataset with BREP-face ground truth and a ParSeNet decomposition baseline.
Key results
- Instant-NGP-derived meshes were non-manifold for nearly 50% of F360 shapes even under idealized conditions.
- Point2CAD outputs were non-manifold for 81% of F360 samples and 92.1% of ABC subset, and 93.1% of Thang3D (6.9% manifold).
- ParSeNet with normals reached matched mIoU 0.23 overall (0.62 without normals, NN) on F360 decomposition with 8.9 false segments on average.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Evaluate Point2CAD on Fusion 360 instead of ABC | Point2CAD · manifold reconstruction | manifold shapes (%): 92.1 -> 18.9 | Hurt |
| Evaluate Point2CAD on Thang3D instead of ABC | Point2CAD · manifold reconstruction | manifold shapes (%): 92.1 -> 6.9 | Hurt |
| Evaluate Point2Cyl on Fusion 360 instead of ABC | Point2Cyl · manifold reconstruction | manifold shapes (%): 81.2 -> 12.3 | Hurt |
| Evaluate Point2Cyl on Thang3D instead of ABC | Point2Cyl · manifold reconstruction | manifold shapes (%): 81.2 -> 0 | Hurt |
| Use ground-truth BREP-face segmentation for reconstruction | Point2CAD on F360 · F360 segmentation dataset | non-manifold reconstruction outputs (%): 81% -> 47% | Helped |
| Oversample points near high-curvature edges | ParSeNet · F360 segmentation dataset | matched mean IoU | Helped |
| Remove surface normals from the input (ablation) | ParSeNet · F360 segmentation dataset, All | matched mean IoU: 0.23 -> 0.62 | Helped |
Surprising findings
- ParSeNet performed worse when surface normals were added than without them, contrary to the original ParSeNet paper.
- Even NeRF-derived meshes at 256^3 resolution under noiseless, idealized conditions were non-manifold for about half of shapes.
Limitations
- ParSeNet decomposition still yields low mmIoU and many false/missing segments, far from practical thresholds.
- Evaluation uses a limited Thang3D subset of 200 models and relies on mesh approximations of STEP files.
- Largely argumentative position with no new model architecture proposed.
Experimental setup
Datasets: Fusion360 Gallery Segmentation Dataset Thang3D ABC
Benchmarks: Fusion360 Gallery Segmentation Dataset Thang3D
Metrics: non-manifold percentage chamfer distance non-manifold edges matched mean IoU missing and false segments
Baselines: Point2CAD Point2Cyl Instant-NGP CAD-MLLM CAD-SIGNet
Builds on: Instant-NGP Point2CAD ParSeNet DeepCAD
Closest papers by meaning
- CraftsMan3D: High-fidelity Mesh Generation with 3D Native Diffusion and Interactive Geometry Refiner (CVPR 2025 · similarity 0.87)
- Generating CAD Code with Vision-Language Models for 3D Designs (ICLR 2025 · similarity 0.86)
- PartGen: Part-level 3D Generation and Reconstruction with Multi-view Diffusion Models (CVPR 2025 · similarity 0.86)
- Interactive3D: Create What You Want by Interactive 3D Generation (CVPR 2024 · similarity 0.85)
- AutoPartGen: Autoregressive 3D Part Generation and Discovery (NeurIPS 2025 · similarity 0.85)
- MeshAnything: Artist-Created Mesh Generation with Autoregressive Transformers (ICLR 2025 · similarity 0.85)
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.