Hi3D v1.0 was released on October 23 with public model files and inference code. It refines a coarse three-dimensional mesh using a reference image of the same object. This is a geometry-refinement stage, rather than evidence that a single photograph becomes an accurately measured digital twin.
The published repository excludes training code and the research evaluation harness. The package therefore enables inference experiments, but not complete reproduction of training and every comparative result. This boundary matters more than a general description of the project as open.
What higher resolution means
The documentation describes successive passes using sparse coordinates, up to a setting labelled 1536³. This does not promise a fixed triangle count or a dense array of those dimensions. The initial mesh supplies structure, while unseen surfaces receive plausible completion rather than measurement.
That distinction changes evaluation. For an artistic asset, a convincing silhouette may be sufficient after editorial review. A part that must fit an assembly requires separate dimensional checks. Additional visible detail does not automatically provide more reliable information about a physical object.
The output is geometry without finished materials and textures. The documented procedure normalises the object's scale. A downstream stage must therefore account for sizing, orientation and integration into the target scene. Those are separate tasks from generating a denser surface.
A local pilot, without a promise for every GPU
The model card specifies Linux and an NVIDIA GPU; the authors validated the highest setting on a 96 GB H20. Minimum memory requirements for smaller cards have not been established. We do not infer such a guarantee from the availability of a mechanism that offloads part of the model to the CPU.
For a pilot, we recommend a small set of the team's own objects with varied shapes: thin parts, recesses and repeating relief. Retain the input image, coarse mesh, version and settings. Compare corrected details, newly introduced errors and the time needed for subsequent manual work. This is a proposed editorial evaluation method; SIExpo has not run the model.
An evaluator should also inspect views absent from the reference image. Otherwise, an attractive front view can hide a problem revealed by rotation, printing or simulation. Acceptance should follow the intended use, with artistic and measurement requirements recorded separately.
Public files carry specific terms
The project uses the Hi3D Open License Agreement 1.0, rather than a standard Apache or MIT licence. It contains conditions covering commercial use, distribution and scale, while external models retain their own terms. Public download should not be described as permission for every conceivable use.
The practical contribution of this release is that teams can examine one stage of a 3D workflow using their own data. Published demonstrations and company claims do not replace their own quality checks. We have no independent benchmark establishing production readiness or superiority over every alternative.
*Cover: a conceptual AI illustration, not output from a Hi3D test.*
*Sources checked: 2026-10-23T16:10:40.049366+00:00.*
