GeoFusion-CAD: Structure-Aware Diffusion with Geometric State Space for Parametric 3D Design
TLDR
GeoFusion-CAD uses a diffusion framework with state-space modeling for scalable, structure-aware parametric CAD generation, outperforming transformers on long sequences.
Reasoning
The paper addresses the critical limitation of long command sequences in parametric CAD generation by introducing a novel diffusion-based approach with hierarchical tree encoding and a C-Mamba block for long-range dependencies. It also contributes a new benchmark, DeepCAD-240, for evaluating long sequences. However, the abstract lacks details on real-world deployment or comparisons beyond the benchmark, and the method's reliance on sketch-extrusion semantics may limit generality.
Read-first score
Read-first score 64.7, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 82.
Field roles
Rank sensitivity
Stability: volatile; rank range: 7.
Keyword Scores
Deep Analysis
Innovations
- Hierarchical tree encoding of CAD programs to jointly capture geometry and topology
- State-space diffusion process with C-Mamba block for long-range structural dependencies via selective state transitions
- Introduction of DeepCAD-240 benchmark for long-sequence evaluation (sequence lengths 40 to 240)
- End-to-end diffusion framework for scalable and structure-aware parametric CAD generation
Methodology
GeoFusion-CAD encodes CAD programs as hierarchical trees and uses a state-space diffusion model with a C-Mamba block to capture long-range dependencies. The model is evaluated on the new DeepCAD-240 benchmark, which extends sequence lengths from 40 to 240 while preserving sketch-extrusion semantics from the ABC dataset.
Key Results
GeoFusion-CAD achieves superior performance on both short and long command sequences, maintaining high geometric fidelity and topological consistency, and sets new state-of-the-art for long-sequence parametric CAD generation where Transformer-based models degrade.