432 points • 1 day ago • Article Link

GenCAD 是 MIT 研究人员开发的一种新型 AI 系统，能够从二维图像生成完整的参数化 CAD 模型。与以往只输出网格或点云等简化表示的方法不同，GenCAD 直接产出可在工程软件中执行的 CAD 命令序列，不仅重建三维形体，还保留了完整的设计命令历史，便于工程师修改与二次编辑。

该系统解决了 AI 驱动设计中的一大难题：传统 CAD 的数据结构（边界表示，B-rep）复杂难以直接被 AI 处理，现有方法常以牺牲精度和可编辑性为代价。 GenCAD 通过以参数化 CAD 命令序列为工作对象——也就是在 CAD 软件中构建模型的逐步指令——规避了这一瓶颈。

其架构由四个关键模块协同工作：首先，自回归 Transformer 编码器将 CAD 命令序列压缩表示到潜在空间；其次，对比学习模型学习图像与 CAD 命令的联合表示，搭建两种模态之间的桥梁；第三，潜在扩散模型根据输入图像生成新的 CAD 潜在表示；最后，解码器将这些潜在表示还原为可被几何内核执行的 CAD 命令，从而生成三维实体模型。

除从图像生成 CAD 外，GenCAD 还可用于 CAD 检索：在数千个模型的数据库中找到与目标图像匹配的现有 CAD 程序。系统还能生成多样化样本——对同一输入图像提供多种不同的 CAD 解释，为设计师提供多个备选方案，从而显著加速设计流程，使工程师能从概念图快速得到可编辑、可用于制造的 CAD 模型。

研究人员认为 GenCAD 是向从图像实现更精确、可修改的三维建模迈出的重要一步。通过保留 CAD 模型的完整参数化历史，该方法保全了工程应用中关键的精度与可编辑性，具有在自动化设计流程、快速原型制作和设计空间探索等方面的潜在应用价值，可简化从概念图像到功能性、可编辑 CAD 模型的转化过程。 GenCAD is a new AI system developed by researchers at MIT that generates complete, parameterized CAD models from 2D images. Unlike previous approaches that produce simplified representations like meshes or point clouds, GenCAD outputs actual CAD command sequences that can be directly used in engineering software. This means it doesn't just create a 3D shape, but generates the entire history of design commands that engineers can modify and work with.

The system addresses a major challenge in AI-driven design. Traditional CAD data structures, known as boundary representation or B-rep, are complex and difficult for AI models to work with directly. Most existing methods sacrifice the precision and editability that make CAD models valuable for manufacturing and design exploration. GenCAD overcomes this by working with parametric CAD command sequences, which are essentially the step-by-step instructions used to build a 3D model in CAD software.

The architecture combines four key components working together. First, an autoregressive transformer encoder learns to represent CAD command sequences in a compressed latent space. Second, a contrastive learning model bridges the gap between CAD commands and images by learning joint representations of both modalities. Third, a latent diffusion model generates new CAD representations based on input images. Finally, a decoder converts these latent representations back into actual CAD commands that can be executed by geometry kernels to produce 3D solid models.

Beyond just generating CAD from images, GenCAD can also perform CAD retrieval, finding existing CAD programs that match a given image from a database of thousands of models. The system demonstrates sample diversity, meaning it can produce multiple different CAD interpretations of the same input image, giving designers various options to choose from. This capability could significantly speed up the design process by allowing engineers to start with an image concept and quickly get editable, manufacturing-ready CAD models.

The researchers see GenCAD as an important step toward more precise and modifiable 3D modeling from images. By preserving the full parametric history of CAD models, their approach maintains the accuracy and editability that are critical for real-world engineering tasks. The work has potential applications in automated design processes, rapid prototyping, and design space exploration, making it easier to go from conceptual images to functional, editable CAD models.