Vibe_CADing

Vibe_CADing is a research repository for conditional CAD generation and retrieval in text-to-CAD pipelines. Its core idea is geometry as code: an LLM writes the shape as a program, the program runs, and errors go back to the LLM for repair. The design adds voxel-based 3D generators, an autoregressive PixelCNN-style 2D CAD layout generator, retrieval-augmented generation over open-source CAD libraries, geometric feedback and manufacturability scoring, and LLM-powered multi-agent orchestration with Planner, Retriever, Generator and Critic roles.

What runs today, per the README, is the LLM-to-code path: CadQuery and build123d harnesses that produce B-rep and STEP output with execute-and-retry loops, a GLSL signed distance function variant with GPU-rendered implicit surfaces, a JSON-spec-to-pythonOCC B-rep path for simple primitives, text-to-3D character animation, and supporting demos for text-to-image-to-SVG, PixelCNN 2D CAD grids and CLIP text and mesh retrieval.

The README's roadmap lists training the diffusion, VAE, VQ-VAE and PixelCNN models, implementing the geometric and structural feedback scoring, building the multi-agent loop, adding text and sketch conditioning, and a GUI-based interactive design agent, as future work. The requirements file is currently empty, with per-module dependencies listed in the README instead.