Previously called Sparc3D, the model refines a coarse 3D mesh using a reference image of the same object, producing geometry-only mesh output through the project's Sparc3DPipeline API.
The release supports 512, 1024, and 1536 cubed extraction-grid settings. The model card says a 512 request runs one refinement pass, 1024 runs two, and 1536 runs three. It describes training on approximately two million curated 3D assets. The repository supplies eight JSON and Safetensors runtime files; the DINOv2 and CLIP conditioning encoders are downloaded separately.
Inference requires Linux, Python 3.10, CUDA 12.0 to 12.5, and an NVIDIA GPU with compute capability 8.0 or newer. There is no CPU inference path. The 1536 setting was validated only on a 96 GB NVIDIA H20, and feasibility or peak memory use on smaller GPUs has not been established.
The release includes inference artifacts and code, but not training code or the paper's evaluation harness. Its model card presents visual comparisons, but does not establish independent performance, production reliability, or a minimum supported GPU-memory requirement.