Media365
The new method could “effectively support the generation of complex geometric structures and reduce segmentation error to an extremely low level” compared with existing technology, according to a paper published on Wednesday on open-access repository arXiv.
The authors – including DeepSeek researcher Liu Wen, Tencent’s Zhao Zibo, HKU professor Ma Yi and Beihang University student Qi Dacheng – have open-sourced the novel approach, with code available at GitHub, the world’s largest developer community.
CAD workflows typically begin with 2D sketches, such as lines and circles, which are then transformed into 3D modelling operations. Existing AI methods that generate 3D objects either consume too many tokens or do not support entity selection, limiting complex editing.
