Give a rough 3D mesh more detail: Hi3D releases refinement weights
3-line summary
- The Hi3D team released v1.0 model weights and inference code on October 23.
- It refines an existing coarse mesh using a reference image of the same object.
- Linux and an NVIDIA GPU are required; the highest resolution was validated only on a 96 GB H20.
The refinement stage of a rough 3D asset can now be tried separately. The Hi3D team released v1.0 weights and inference code on October 23. Unlike its October 16 preview, which contained no implementation or weights, the Hugging Face repository now lists actual weight files.
Built on Sparc3D, the model takes a coarse mesh—a surface represented by connected polygons—and a reference image of the same object. It produces geometry without textures, materials or colors. Major structural errors can survive refinement, and unseen surfaces are plausible completions.
The installation guide requires Linux and an NVIDIA GPU, with no CPU inference. The highest-resolution path was validated only on a 96 GB NVIDIA H20. Feasibility and peak memory on smaller cards remain unestablished; 96 GB is not a stated minimum requirement.
This is an inference release. Training code, the evaluation harness and benchmark object lists needed to reproduce the paper’s comparison tables are absent. Hi3D-owned code and weights are covered by a custom license; review its scope and commercial terms before integrating the model into a service.
Why it matters
AI Ping sees another option for evaluating coarse creation and geometric refinement separately. Production suitability still needs checks across multiple views and the amount of editing required afterward.
Who it's for
- Designers
- Pair an existing coarse asset with its reference image, then inspect hidden surfaces and remaining texturing work.
- Developers
- Check the GPU setup and license, and distinguish running inference from reproducing the paper’s comparisons.
