Research map / 3D vision / 3D generation
CAGE: Controllable Articulation 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.
Introduces CAGE, a DDPM-based diffusion model with attribute attention blocks that generates articulated 3D objects from a category label and part connectivity graph, with part retrieval to build meshes.
Problem
Controllable generation of 3D articulated objects, where existing generative models (e.g., NAP) often ignore input constraints on graph structure, part attributes, and articulation.
Approach
A denoising diffusion model over shape-abstraction and articulation attributes per part, with three attention modules (local, global, and graph-relation attention masked by the input adjacency matrix) inside adaLN-Zero attribute attention blocks, enabling fine-grained conditioning on graphs, categories, and part attributes.
Key results
- On K=32 setting, CAGE achieves MMD-AID 0.816 and MMD-ID 0.049 vs NAP-large's 0.860 and 0.067, with AOR 0.008 vs 0.034.
- Human study with 44 participants preferred CAGE outputs 82.61% of the time over NAP-large (17.39%).
- Ablations show removing graph relation attention raises MMD-ID from 0.049 to 0.157.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CAGE's shape abstraction and attention-based diffusion model instead of NAP | Articulated object generation with K=8 · PartNet-Mobility test set (K=8) | MMD (ID): 0.060 -> 0.043 | Helped |
| Use CAGE instead of NAP-large | Articulated object generation with K=32 · PartNet-Mobility test set (K=32) | AOR: 0.034 -> 0.008 | Helped |
| Use CAGE instead of NAP-large | Articulated object generation with K=32 · PartNet-Mobility test set (K=32) | Human preference (%): 17.39 -> 82.61 | Helped |
| Add global attention (GA) to the full model (ablation) | CAGE · PartNet-Mobility test set (K=32) | MMD (ID): 0.053 -> 0.049 | Helped |
| Add global attention (GA) to the full model (ablation) | CAGE · PartNet-Mobility test set (K=32) | AOR: 0.016 -> 0.008 | Helped |
| Add graph relation attention (GRA) to the full model (ablation) | CAGE · PartNet-Mobility test set (K=32) | MMD (ID): 0.157 -> 0.049 | Helped |
| Add graph relation attention (GRA) to the full model (ablation) | CAGE · PartNet-Mobility test set (K=32) | AOR: 0.013 -> 0.008 | Helped |
| Add local attention (LA) to the full model (ablation) | CAGE · PartNet-Mobility test set (K=32) | MMD (ID): 0.052 -> 0.049 | Helped |
| Add local attention (LA) to the full model (ablation) | CAGE · PartNet-Mobility test set (K=32) | AOR: 0.011 -> 0.008 | Helped |
Surprising findings
- Removing local attention or global attention yields only moderate degradation, while the graph relation attention is the dominant module for conforming to graph structure.
Limitations
- Data imbalance leads to better performance on high-frequency object categories.
- Motion attributes show weaker controllability than geometric attributes.
- Part synthesis relies on retrieval from training parts, limiting diversity; overlapping parts and collisions occur.
Experimental setup
Datasets: PartNet-Mobility
Benchmarks: PartNet-Mobility
Metrics: MMD COV 1-NNA ID AID AOR human preference rate
Baselines: NAP NAP-light NAP-large
Builds on: DDPM NAP adaLN-Zero DiT
Closest papers by meaning
- SINGAPO: Single Image Controlled Generation of Articulated Parts in Objects (ICLR 2025 · similarity 0.91)
- MeshArt: Generating Articulated Meshes with Structure-Guided Transformers (CVPR 2025 · similarity 0.90)
- ArtFormer: Controllable Generation of Diverse 3D Articulated Objects (CVPR 2025 · similarity 0.90)
- MIDGArD: Modular Interpretable Diffusion over Graphs for Articulated Designs (NeurIPS 2024 · similarity 0.89)
- DIPO: Dual-State Images Controlled Articulated Object Generation Powered by Diverse Data (NeurIPS 2025 · similarity 0.89)
- General Articulated Objects Manipulation in Real Images via Part-Aware Diffusion Process (NeurIPS 2024 · similarity 0.88)
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.