eesjgong/graph-cad

Graph-CAD is a research project that lets you describe a 3D object in plain English and get back working Blender code that builds it, a process called Text-to-CAD. Instead of jumping straight from your words to executable code (which tends to produce cascading errors in complex models), Graph-CAD first builds an intermediate "decomposition graph", a structured map of the object's parts and how they fit together geometrically, and then converts that map into code in two more steps. The pipeline works in three stages. First, a model reads your text instruction and predicts a hierarchical graph that breaks the object down into sub-components, capturing spatial relationships like alignment and attachment. Second, a planning model converts that graph into an ordered sequence of CAD actions. Third, a code generation model turns those actions into executable Blender Python code you can actually run. To make the system more robust on complicated assemblies, the project uses a training technique called progressive curriculum learning, which automatically creates increasingly difficult training examples near the edge of what the model can handle, pushing its capability boundary outward over time. The project also introduces BlendGeo, a 12,000-example dataset pairing text instructions with decomposition graphs, action sequences, and runnable Blender code. The paper was accepted at ICLR 2026. Pre-trained model weights are available on ModelScope. The full README is longer than what was provided.