HiFi-BRep: High-Fidelity Latent Representation for Robust B-Rep Generation

The Core Problem: B-Rep Models Keep Breaking

Boundary representation (B-Rep) is how CAD software stores 3D shapes—as networks of faces, edges, and vertices that define solid objects. Deep learning has started tackling B-Rep generation, but current methods fail in two critical ways. First, their learned representations are noisy and contaminated, making them unstable starting points for generation. Second, the generation itself cascades errors: predicting one piece of geometry influences the next, and enforcing that shapes are valid (a non-differentiable constraint) doesn't play well with neural networks optimized via gradient descent. The result is high-fidelity geometry that's structurally invalid, or valid shapes that look wrong.

How HiFi-BRep Fixes It

The authors propose two complementary fixes, each tackling one brittleness problem.

Cleaner latent codes. Standard encoders pad variable-length shape data with zeros to fit fixed tensors, and this padding noise leaks into learned features. HiFi-BRep replaces padding with learnable queries—the encoder learns to extract just the shape-relevant information without dummy values. To prevent different parts of the shape from contaminating each other's representations, the encoder uses topology-guided attention, where the attention mechanism respects the actual connectivity of edges and faces. Think of it as teaching the encoder to look only at nearby, related components rather than treating the entire shape as a uniform blob.

Geometry and topology together. Instead of predicting shape components sequentially (which causes errors to compound), HiFi-BRep uses a single-stage decoder that predicts all geometry and topology in parallel. Critically, it embeds manifold constraints—the mathematical rules that make a valid 3D surface—as a differentiable loss function rather than enforcing them after-the-fact with non-differentiable checks. This lets gradients flow through the validity constraint during training, so the model learns to generate valid shapes naturally rather than fighting against the geometry it just created.

Why This Matters

The approach is concrete and addresses root causes. Current methods fail because they treat representation learning and generation as separate problems, and because they use discrete validity checks that don't integrate with gradient-based optimization. HiFi-BRep unifies these: cleaner representations feed into a decoder that is itself constrained to produce valid outputs.

Experimentally, the authors report significant improvements over state-of-the-art baselines in both structural validity (the generated shape actually satisfies B-Rep rules) and geometric fidelity (the shape matches the intended design). For CAD applications—where a single invalid edge or face breaks downstream tools—this robustness is essential. The code is open-source, lowering the barrier for adoption and reproduction.

The insight is pragmatic: instead of trying to generate first and patch validity later, embed validity into the generative process itself through differentiable constraints and parallel prediction. This is a clean methodological contribution that should generalize beyond B-Reps to other structured 3D synthesis tasks.