AI product Open source
Hi3D v1.0 is an AI tool for refining a coarse 3D mesh into higher-resolution geometry conditioned on a reference image of the same object. Built on the Sparc3D representation and refinement architecture, it extracts sparse coordinates from the coarse geometry, conditions sparse flow with image features, progressively samples sparse latents, decodes them with a self-pruning sparse VAE, extracts a mesh, and revoxelizes that mesh to provide support for the next refinement pass. It supports nominal resolutions of 512, 1024, and 1536, corresponding to one, two, or three passes, and exposes Python and command-line inference interfaces that accept an image and coarse mesh and export a geometry-only mesh.
The project provides inference code, model artifacts, and documentation, with model files hosted on Hugging Face. It requires Linux, Python 3.10, and an NVIDIA GPU with compute capability 8.0 or newer using a CUDA 12.0–12.5-compatible PyTorch environment; there is no CPU inference path. The repository does not include training code, the evaluation harness, or benchmark data, and the output does not include textures, materials, or vertex colors. It is distributed under the Hi3D Open License Agreement for Hi3D-owned materials, with separate Apache-2.0, MIT, and third-party terms applying to other components.
1 use taken from transcripts — each links to the moment in the video.
A tool that refines 3D mesh geometry by processing a coarse model and a reference image of the same object through progressively finer computational passes.
1 in the library.