HiFi-BRep: High-Fidelity Latent Representation for Robust B-Rep Generation
HiFi-BRep targets the brittleness that makes generated CAD boundary representations invalid or low fidelity.
The paper says current deep B-Rep generators are weakened by noisy padded latent features and by sequential generation errors. Its encoder uses learnable queries and topology-guided attention to build cleaner latent representations. Its decoder predicts geometry and topology together, with manifold constraints used as a differentiable training objective. The authors report better structural validity and geometric fidelity than prior methods, and say code and models are public. HF Daily Papers' note
The paper says current deep B-Rep generators are weakened by noisy padded latent features and by sequential generation errors. Its encoder uses learnable queries and topology-guided attention to build cleaner latent representations. Its decoder predicts geometry and topology together, with manifold constraints used as a differentiable training objective. The authors report better structural validity and geometric fidelity than prior methods, and say code and models are public. HF Daily Papers' note
score 4