Home  >  Software

October 24, 2026      Software      3DPTimes Editorial | Sources: Hi3D announcement, October 23, 2026; official GitHub and Hugging Face release documentation.      21

The Chinese AI 3D developer’s October 23 release makes image-guided mesh refinement available under a custom license, with clear limits for printing workflows.

Archive photograph: a physical 3D-printed object optimized with netfabb, photographed July 1, 2010. It is not a Hi3D-generated model, a Hi3D product or a demonstration of this release. Photo: Creative Tools, CC BY 2.0. Original: https://commons.wikimedia.org/wiki/File:3D_printed_object_made_with_netfabb_(4751093079).jpg | License: https://creativecommons.org/licenses/by/2.0/ . Reduced resolution; image content unchanged.

Hi3D has released inference code and model weights for Hi3D V1.0, opening its geometry-refinement stage to developers and researchers. The Beijing company announced the release on October 23, with files available through its official GitHub and Hugging Face repositories.

The package takes an existing coarse mesh and a reference image of the same object, then refines the geometry. It builds on Sparc3D. Users need an upstream method to create the initial mesh; the released refinement component does not generate textures or materials.

The model documentation describes training on approximately two million curated 3D assets and an effective sparse extraction grid reaching 1536 cubed. That figure describes the extraction grid, rather than a fixed polygon count or guaranteed printable feature size.

For additive manufacturing, access to the refinement stage creates another option between an initial AI-generated shape and downstream mesh preparation. Hi3D reports a 90% slicing-success rate in its production workflow, but this is a company-reported slicing result, not an independently verified physical printing-success rate.

The repository includes inference code, but does not provide training code or the evaluation harness needed to reproduce the paper’s comparative results. Its highest-resolution configuration has been validated on a 96 GB NVIDIA H20; feasibility on smaller GPUs has not been established. Output geometry is normalized, so users must set the intended physical dimensions before printing.

The release uses the Hi3D Open License Agreement Version 1.0. Commercial use is permitted subject to its conditions, including branding requirements for commercial integrations and separate licensing procedures above specified user or revenue thresholds. Third-party components retain their own terms. Developers should review the agreement for their deployment.

Original sources: Hi3D company announcement, October 23, 2026: https://www.prnewswire.com/news-releases/hi3d-v1-0-goes-open-source-the-first-production-validated-open-source-ai-3d-model-302917973.html
Official release and technical documentation, released October 23, checked October 24: https://github.com/hi3d-official/Hi3Dv1.0
Model card and weights: https://huggingface.co/hi3d-official/Hi3Dv1.0
Release license: https://github.com/hi3d-official/Hi3Dv1.0/blob/main/LICENSE







©2026 3dptimes.com All Rights Reserved