Beyond Spatial Compression: Interface-Centric Generative States for Open-World 3D Structure
C2LT-3D introduces interface-centric generative states to improve assembly robustness in open-world 3D structures by separating geometry, ownership, and seam relations.
Key Findings
Methodology
C2LT-3D separates 3D representation into local geometry, partition-conditioned context, and seam variables. It uses unsupervised partitioning and seam priors to support assembly validation, repair, and constrained decoding.
Key Results
- On Objaverse-LVIS, C2LT-3D achieves a separation score of 0.9780, outperforming BPT's 0.9220, with a contamination rate of just 0.0141, significantly better than baselines.
- In zero-shot transfer, C2LT-3D achieves Chamfer distance 0.0268 and Hausdorff distance 0.2282, surpassing existing methods.
- Seam repair experiments show C2LT-3D is more robust under adversarial assembly scenarios, reducing structural errors significantly.
Significance
This work redefines evaluation criteria for 3D generative representations, emphasizing operational discrete states for assembly-level reasoning, addressing challenges in complex 3D asset generation and repair.
Technical Contribution
C2LT-3D introduces interface-centric generative states, addressing the entanglement of local shape, ownership, and assembly relations in compression-based representations, offering a new framework for open-world 3D generation.
Novelty
This is the first method to propose interface-centric generative states, transforming 3D generation from passive compression to active structural construction, enabling robust assembly and repair.
Limitations
- Relies on the accuracy of unsupervised partitioning, which may introduce errors in complex components.
- Training the seam prior requires significant computational resources.
- Does not fully address global topological closure issues.
Future Work
Future work could explore more efficient partitioning algorithms, stronger global topological constraints, and applications in dynamic scenarios.
AI Executive Summary
Current 3D generation methods focus on spatial compression, neglecting explicit modeling of component ownership and assembly relations, leading to poor performance in open-world assets.
This paper introduces C2LT-3D, which uses interface-centric generative states to separate 3D representation into local geometry, partition-conditioned context, and seam variables. It supports assembly validation, structural repair, and constrained decoding without requiring post-hoc recovery modules.
Experiments demonstrate that C2LT-3D outperforms baselines on Objaverse-LVIS, achieving superior robustness in zero-shot transfer and adversarial assembly scenarios. While promising, challenges like global topological closure remain open for future exploration.
Deep Analysis
Background
Traditional 3D generation methods like PolyGen and MeshGPT focus on compressing geometry, performing well on single-component CAD models but failing on complex multi-component assets.
Core Problem
Existing methods entangle local shape, ownership, and assembly relations in latent streams, lacking support for assembly-level reasoning, leading to frequent structural errors in open-world assets.
Innovation
C2LT-3D introduces interface-centric generative states, decomposing representation into local geometry, partition-conditioned context, and seam variables, addressing three major failure modes of compression-based tokens.
Methodology
- �� Local geometry encoding: Partial canonicalization removes pose variation.
- �� Partition-conditioned context: Unsupervised partitions reduce cross-component interference.
- �� Seam prior: Predicts seam compatibility, transforms, and collision risks.
Experiments
Trained on ShapeNet, tested zero-shot on Objaverse-LVIS. Compared against BPT and VQ-Patch using Chamfer distance, separation scores, and other metrics.
Results
C2LT-3D outperforms baselines in separation scores, contamination rates, and Chamfer distance, excelling in complex assembly scenarios.
Applications
Applicable for generating, repairing, and validating assemblies in complex 3D assets, such as industrial design and virtual reality.
Limitations & Outlook
Relies on partition and seam prediction accuracy, does not fully resolve global topology, and has high computational costs.
Plain Language Accessible to non-experts
Imagine a jigsaw puzzle where each piece has shape, color, and edge details. Traditional methods focus only on shape, while C2LT-3D also tracks which piece belongs to which section and which edges can connect. This makes assembly and repair much easier.
ELI14 Explained like you're 14
Think of playing with LEGO bricks. Regular methods just tell you the shape of each brick. C2LT-3D tells you which set each brick belongs to and checks if they can snap together. Cool, right?
Glossary
C2LT-3D
An interface-centric 3D generation method that separates geometry, context, and seam variables.
Used for open-world 3D asset generation and repair.
Interface-centric generative state
A representation explicitly modeling geometry, ownership, and assembly relations.
Core concept of C2LT-3D.
Partition-conditioned context
Context modeling that reduces interference between components using unsupervised partitions.
Supports component-level reasoning.
Seam prior
A model predicting seam compatibility, transforms, and collision risks.
Enables assembly validation and repair.
Objaverse-LVIS
A dataset of complex multi-component assets for open-world scenarios.
Used for zero-shot testing.
Open Questions Unanswered questions from this research
- 1 How to improve unsupervised partition accuracy to reduce errors?
- 2 How to efficiently achieve global topological closure?
- 3 Can this method extend to dynamic scenarios?
Applications
Immediate Applications
Industrial Design
Supports assembly validation and repair for complex mechanical components, improving design efficiency.
Virtual Reality
Generates and interacts with multi-component objects in virtual environments.
Long-term Vision
Dynamic Scene Modeling
Extends to dynamic environments, enabling real-time assembly and repair.
Abstract
Current 3D tokenizers largely treat representation as spatial compression: compact codes reconstruct surface geometry, but leave component ownership and attachment validity implicit. In open-world assets with intersecting components, noisy topology, and weak canonical structure, this creates a representation mismatch: local shape, component identity, and assembly relations become entangled in a latent stream and are not natively addressable during decoding. We formulate an alternative view, interface-centric generative states, in which tokenization constructs an operational state rather than a passive compressed code. The state exposes local geometry, component ownership, and attachment validity as variables that can be queried, constrained, and repaired during decoding. We instantiate this formulation with Component-Conditioned Canonical Local Tokens (C2LT-3D), factorizing representation into canonical local geometry, partition-conditioned context, and relational seam variables. Each factor targets a distinct failure mode of compression-centric tokens: pose leakage, cross-component interference, or invalid local attachment. This exposed state supports attachment validation, latent structural repair, targeted intervention, and constrained serialization without a separate post-hoc structure recovery module. Trained on single-object CAD models and evaluated zero-shot on open-world multi-component assets, C2LT-3D improves structural robustness and shows that its latent variables remain actionable under adversarial attachment settings. These results suggest that open-world 3D generative representations should be evaluated not only by reconstruction fidelity, but by whether their discrete states remain operational for assembly-level structural reasoning.