Show HN: Code3DBench – single-image to executable Three.js benchmark (github.com)

Code3DBench has been introduced as a novel benchmark for evaluating multimodal models that generate executable Three.js code from a single rendered image of a low-poly 3D object. This initiative allows researchers to assess the models’ abilities to reconstruct objects inside a predefined browser environment. The benchmark includes a comprehensive repository featuring a prompt template, a fixed Three.js scaffold, and a dataset of 1,012 public CC0 low-poly objects spanning eight categories such as vehicles, characters, and furniture. The generated code is evaluated based on executability, mesh geometry quality, and image-space diagnostics using metrics like Chamfer distance and structural similarity indices. This development is significant for the AI/ML community as it lays the groundwork for improving the automated generation of 3D content, a crucial area in gaming, virtual reality, and online visualization. The ability to translate a single image into functional code can dramatically streamline workflows for developers and artists, making 3D modeling more accessible. Furthermore, the results highlight that while runtime success for generated code improves rapidly, challenges remain with exported mesh geometry, pointing to areas for further research and enhancement within AI-driven 3D generation.