wtf is light-sq?

Light-SQ is the code release for a research paper presented at SIGGRAPH Asia 2025. It tackles a problem in 3D graphics called shape abstraction: taking a detailed 3D mesh, the kind produced by 3D generation tools, and describing its overall form using a small number of simple rounded box like shapes called superquadrics. The goal is to capture the structure of an object, such as its major parts and how they fit together, rather than every small surface detail. What makes this project distinct is that it pays close attention to the underlying structure of the object while doing this simplification, instead of just approximating the outer surface. It offers two ways to run the fitting process: one that works directly from the mesh, and one that first breaks the mesh into structurally meaningful convex pieces before fitting superquadrics to each piece, which tends to produce a more structurally faithful result. Using it involves several command line steps run in sequence: first converting an input mesh into a signed distance field representation on a GPU, then running a fitting script that produces the final superquadric shapes, with an optional extra step that adds the structure aware decomposition guidance. The output is saved to a folder alongside visualizations if requested. This is research code aimed at people already working in 3D graphics, computer vision, or generative 3D modeling. It requires an NVIDIA GPU with CUDA, a specific Python environment built with conda, and several exact library versions including PyTorch, Open3D, and a custom optimization library bundled in the repository. Some pieces mentioned in the project's todo list, like faster parallel processing and a code release for one component, are not yet available. The authors ask that anyone using the results in their own research cite the accompanying paper.