TL;DR
AutoBrep is a single autoregressive Transformer that generates Computer-Aided Design (CAD) solid models by encoding both geometry and topology into a unified sequence of discrete tokens. By framing 3D modeling as a next-token prediction task following a breadth-first traversal, it eliminates the need for complex multi-stage pipelines, delivering a 2× to 5× inference speedup and significantly higher topological validity and scalability than prior state-of-the-art methods.
What Problem Does This Paper Address?
In CAD, the gold standard for representing 3D shapes is the Boundary Representation (B-Rep). A B-Rep model consists of two tightly coupled components: continuous geometry (the mathematical surfaces and curves) and discrete topology (the graph structure defining how faces, edges, and vertices connect). For a B-Rep to be valid, it must be "watertight"—meaning all pieces must align with absolute mathematical precision.
Directly generating B-Reps using deep learning is notoriously difficult. Previous approaches relied on multi-stage pipelines (e.g., separate neural networks to predict vertices, then edges, then faces, followed by multiple diffusion models to generate geometry). These cascaded systems suffer from cumulative error, struggle to learn the joint distribution of geometry and topology, and fail to scale to complex shapes, typically capping out at models with fewer than 50 faces.
Methods
AutoBrep solves this by unifying geometry and topology into a single stream of discrete tokens, allowing a standard GPT-style Transformer to handle the entire generation process end-to-end.
1. Geometry Tokenization
To represent arbitrary parametric surfaces and curves uniformly, the authors sample them into point grids (32×32 for faces, 32×1 for edges). They pass these grids through a Deep Compression Autoencoder paired with Finite Scalar Quantization (FSQ) to compress the continuous coordinates into compact integer tokens (4 tokens per face, 2 tokens per edge). They also introduce a simple coordinate-sorting normalization step to make the representation invariant to how the underlying UV-space was parameterized.
2. Topology Tokenization via Breadth-First Traversal (BFT)
Instead of using a global coordinate ordering, AutoBrep traverses the B-Rep's face adjacency graph using a breadth-first search. This traversal orders the sequence level by level.
Crucially, because of the properties of BFT, an edge at level $l$ can only connect to faces in the immediate previous level ($l-1$) or within the same level ($l$). This allows the authors to implement a Local Topological Reference scheme. Instead of using global face IDs, the model uses local reference tokens ($T$) that only look back at a small moving window of the two most recent BFT levels. This keeps the vocabulary size small and bounded.
3. Autocomplete Fine-Tuning
Because the model treats generation as a sequence, B-Rep autocompletion becomes a natural extension. User-provided faces are simply injected as the first level of the BFT sequence. The model is then fine-tuned to predict the remaining levels, guaranteeing that the user's input geometry is perfectly preserved.
Datasets
The authors introduced two large-scale datasets derived from the open-source ABC dataset:
- ABC-1M: Contains ~1.3 million unique, single-body CAD models decomposed from multi-body STEP files, filtered to remove duplicates and non-manifold geometry.
- ABC-Constraint: A dataset specifically designed for autocompletion. It identifies common engineering interfaces—like cylindrical bolt holes and planar faces on a part's convex hull—to simulate realistic user inputs.
Key Insights and Ideas
- CAD Generation as Language Modeling: By discretizing both the continuous geometry and the discrete graph structure, B-Rep generation maps cleanly onto standard autoregressive language modeling architectures.
- The Power of Locality: Using a breadth-first traversal reveals a localized structure in CAD topology. Exploiting this with a moving context window prevents the model from needing global tracking, which drastically improves generation quality and reduces overfitting on long sequences.
- Inherent Autocompletion: Treating user-guided generation as a conditional prefix sequence avoids the complex masking or post-optimization tricks required by diffusion models, ensuring 100% fidelity to user-specified faces.
Key Results
- Higher Watertight Validity: AutoBrep achieved an overall valid/watertight ratio of 70.8%, outperforming baselines like BrepGen (46.6%) and HoLa (54.8%).
- Unprecedented Scalability: While baseline models quickly degrade to 0% validity as the face count approaches 50, AutoBrep maintains roughly a 50% validity rate for complex solids up to 100 faces.
- Generation Speed: Thanks to KV-caching inherent to autoregressive Transformers, AutoBrep achieves an average inference time of 0.46 seconds per face, making it 2× to 5× faster than diffusion-based alternatives.
- Qualitative Diversity: Visual evaluations show that AutoBrep successfully generates complex mechanical features like chamfers, fillets, and intricate internal topologies that baselines struggle to replicate.
Limitations and Future Work
Despite its strengths, AutoBrep exhibits a few common failure modes:
- Extreme Aspect Ratios: Long, winding, narrow faces—such as screw threads—are poorly captured by uniform point grids.
- Thin Shells & Sliver Faces: Very thin structures occasionally fail to generate correctly due to the precision limits of bounding-box quantization.
- Self-Intersections: In highly crowded volumes, the model can occasionally generate overlapping faces.
For future work, the authors suggest filtering out standard components (like bolts) that are better sourced from catalogs, improving bounding-box precision, and exploring alternative geometric representations beyond uniform point grids to better capture sharp mechanical details.
Summary updated