Research map / Point clouds / Shape and surface completion
Flatten Anything: Unsupervised Neural Surface Parameterization
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 the Flatten Anything Model (FAM), an unsupervised per-model neural network that learns global free-boundary UV parameterization of 3D surface points via a bi-directional cycle mapping of cut, unwrap, deform and wrap sub-networks.
Problem
Global free-boundary surface parameterization of 3D surfaces, which traditionally requires high-quality meshes, manual seam cutting, and connectivity information, and does not extend easily to unstructured point clouds.
Approach
Learn point-wise 3D-to-UV mappings with MLP sub-networks (Deform, Wrap, Cut, Unwrap) assembled into two parallel cycle-consistent branches, so that cutting seams and UV boundaries are learned adaptively from points without connectivity, with no ground-truth UV maps.
Key results
- On 8 open-surface models, FAM achieves lower conformality error than SLIM on all models (e.g., human-face 0.074 vs 0.635, car-shell 0.037 vs 0.411).
- On unstructured point clouds, FAM's conformality (0.037-0.117) is worse than FBCP-PC (0.019-0.023), which the authors attribute partly to FBCP-PC receiving manually specified boundary indices.
- FAM optimization takes around 18 minutes per model on a single RTX 3090, compared with 17-40 minutes for SLIM on a CPU.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FAM instead of SLIM | Surface parameterization · human-face | conformality: 0.635 -> 0.074 | Helped |
| Use FAM instead of SLIM | Surface parameterization · human-head | conformality: 0.254 -> 0.094 | Helped |
| Use FAM instead of SLIM | Surface parameterization · car-shell | conformality: 0.411 -> 0.037 | Helped |
| Use FAM instead of SLIM | Surface parameterization · spiral | conformality: 0.114 -> 0.087 | Helped |
| Use FAM instead of SLIM | Surface parameterization · human-hand | conformality: 0.609 -> 0.145 | Helped |
| Use FAM instead of SLIM | Surface parameterization · shirt | conformality: 0.443 -> 0.166 | Helped |
| Use FAM instead of SLIM | Surface parameterization · three-men | conformality: 0.645 -> 0.162 | Helped |
| Use FAM instead of SLIM | Surface parameterization · camel-head | conformality: 0.349 -> 0.088 | Helped |
| Use FAM instead of FBCP-PC | Point-cloud parameterization · cloth-pts (7K points) | conformality: 0.021 -> 0.037 | Hurt |
| Use FAM instead of FBCP-PC | Point-cloud parameterization · julius-pts (11K points) | conformality: 0.019 -> 0.058 | Hurt |
| Use FAM instead of FBCP-PC | Point-cloud parameterization · spiral-pts (28K points) | conformality: 0.023 -> 0.117 | Hurt |
| Add the 2D→3D→2D cycle branch (ablation) | FAM · Surface parameterization | parameterization quality | Helped |
Surprising findings
- On point clouds, FAM underperforms FBCP-PC in conformality, which the authors note is a less-than-fair comparison because FBCP-PC uses manual boundary input.
- Self-intersections occur in both FAM and SLIM, though at very small proportions.
Limitations
- Per-model overfitting; cannot exploit existing UV results to train a generalizable model.
- Inferior quality on highly complex CAD models with interior structures and multi-layer issues.
- Does not consider shape symmetry, seam visibility, or seamless parameterization.
Experimental setup
Datasets: 3D surface models collected by the authors (e.g., human-face, brain, cow, fandisk, nefertiti, dragon) cloth-pts, julius-pts, spiral-pts point clouds
Benchmarks: 8 open-surface/disk-topology models vs SLIM 16 diverse models with varied topology 3 point clouds vs FBCP-PC
Metrics: conformality (average absolute angle difference) self-intersection proportion optimization time
Baselines: SLIM FBCP-PC Nuvo OptCuts DiffSR
Builds on: FoldingNet AtlasNet geometry images SLIM
Closest papers by meaning
- M-plicits: Neural Implicit Surfaces via Nested Multiscale Residuals (NeurIPS 2026 · similarity 0.87)
- Implicit-ARAP: Efficient Handle-Guided Neural Field Deformation via Local Patch Meshing (NeurIPS 2025 · similarity 0.87)
- 4Deform: Neural Surface Deformation for Robust Shape Interpolation (CVPR 2025 · similarity 0.86)
- Improving Neural Network Surface Processing with Principal Curvatures (NeurIPS 2024 · similarity 0.86)
- Small Steps and Level Sets: Fitting Neural Surface Models with Point Guidance (CVPR 2024 · similarity 0.86)
- High Resolution UDF Meshing via Iterative Networks (NeurIPS 2025 · similarity 0.86)
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.