Research map / 3D vision / Face and shape
ShapeWalk: Compositional Shape Editing Through Language-Guided Chains
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.
Introduces ShapeWalk, a synthetic dataset of 158K CAD-derived shapes linked into 26K language-annotated edit chains built by interpolating GeoCode shape-program parameters, and trains latent neural editors with new multi-step editing metrics.
Problem
Language-guided 3D shape editing lacks large-scale data with precise ground-truth edits and multi-step (compositional) edit chains; existing datasets like ShapeTalk provide shape pairs with many confounded factors of variation.
Approach
Synthesize edit chains by reconstructing realistic CAD shapes (3DCOMPAT++) into the GeoCode shape-program parameter space and interpolating between shape pairs, so each edge has a single controlled parameter edit with exact ground-truth; instructions are produced by rule-based generation plus T5 paraphrasing.
Key results
- Coupled OURS512x4 achieves better Chamfer Distance than coupled LATEFUSION512 (averaged FCD 2.670 vs 3.002 is for a different config; OURS512x4 coupled averaged FCD 2.990 vs LATEFUSION512 2.719 per Table 1), with decoupled direction/magnitude variants generally outperforming coupled ones.
- Edit recognition classifier reaches 96.9% accuracy on the realistic set and 88.4% on the random set, dropping to 58.7% for legs roundness/indentation.
- Directly predicting the edited latent (DIRECTGEN) performs poorly across all chained-editing metrics.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Predict edit direction and magnitude separately instead of a coupled edit vector (ablation) | OURS512×4 · ShapeWalk chains |P| ∈ {10, 15, 20} | FCD×1e4: 2.990 -> 2.670 | Helped |
| Predict edit direction and magnitude separately instead of a coupled edit vector (ablation) | OURS512×4 · ShapeWalk chains |P| ∈ {10, 15, 20} | ACD×1e4: 2.703 -> 2.524 | Helped |
| Use a 4-layer MLP rather than an 8-layer MLP (ablation) | OURS512 · ShapeWalk chains |P| ∈ {10, 15, 20} | FCD×1e4: 3.208 -> 2.990 | Helped |
| Use a 4-layer MLP rather than an 8-layer MLP (ablation) | OURS512 (decoupled) · ShapeWalk chains |P| ∈ {10, 15, 20} | FL2: 1.497 -> 1.447 | Helped |
| Use a 1024-dimensional rather than 512-dimensional bottleneck (ablation) | LATEFUSION (coupled) · ShapeWalk chains |P| ∈ {10, 15, 20} | FCD×1e4: 2.856 -> 2.719 | Hurt |
| Use a 256-dimensional rather than 512-dimensional bottleneck (ablation) | LATEFUSION (coupled) · ShapeWalk chains |P| ∈ {10, 15, 20} | FCD×1e4: 2.719 -> 2.874 | Hurt |
| Use a 256-dimensional rather than 1024-dimensional bottleneck (ablation) | LATEFUSION (decoupled) · ShapeWalk chains |P| ∈ {10, 15, 20} | FCD×1e4: 2.711 -> 3.309 | Hurt |
| Use OURS512×4 rather than LATEFUSION512 (ablation) | Coupled latent mapper · ShapeWalk chains |P| ∈ {10, 15, 20} | FCD×1e4: 2.719 -> 2.990 | Mixed |
| Train the edit classifier on realistic shapes and evaluate on random shapes | Shape edit recognition · Random set | accuracy: 0.884; 0.969 | Mixed |
| Recognize subtle legs roundness/indentation edits rather than high-variation seat height edits | Edit classifier · Shape edit recognition | accuracy: 1.000 -> 0.587 | Hurt |
Surprising findings
- Directly predicting the edited shape latent without an edit vector leads to poor edit predictions across all metrics.
- The edit recognizer struggles on subtle or part-count edits (e.g., adding/removing legs or backrest rails at 66.3%) despite high accuracy on large changes.
Limitations
- Neural editor quality is bounded by the shape autoencoder; PC-AE cannot capture very fine-grained edits.
- Edit recognition fails on subtle changes and on counting-related edits such as adding or removing parts.
- Random set does not necessarily produce realistic shapes; realistic set is limited to 1113 reconstructed shapes.
Experimental setup
Datasets: 3DCOMPAT++ ShapeNet ShapeWalk
Benchmarks: ShapeWalk
Metrics: Chamfer Distance L2 latent distance average edit error final edit error edit recognition accuracy
Baselines: DIRECTGEN LATEFUSION ShapeTalk-style LATEFUSION
Builds on: GeoCode ShapeTalk ChangeIt3D PC-AE
Closest papers by meaning
- CADMorph: Geometry‑Driven Parametric CAD Editing via a Plan–Generate–Verify Loop (NeurIPS 2025 · similarity 0.87)
- Part-X-MLLM: Part-aware 3D Multimodal Large Language Model (ICLR 2026 · similarity 0.87)
- ShapeCraft: LLM Agents for Structured, Textured and Interactive 3D Modeling (NeurIPS 2025 · similarity 0.86)
- CADTalk: An Algorithm and Benchmark for Semantic Commenting of CAD Programs (CVPR 2024 · similarity 0.86)
- ShapeWords: Guiding Text-to-Image Synthesis with 3D Shape-Aware Prompts (CVPR 2025 · similarity 0.86)
- EditRoom: LLM-parameterized Graph Diffusion for Composable 3D Room Layout Editing (ICLR 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.