Hi3D releases open-source Hi3D v1.0, an image-to-3D model

Summary

Hi3D v1.0 (formerly Sparc3D) is an open-source tool for converting images into 3D models. The model refines a coarse mesh using three-pass sparse-voxel operations up to 1536 resolution. Developers claim state-of-the-art multi-view fidelity and a 90% 3D printing slicing success rate. Code and weights are available on GitHub/Hugging Face. Commercial use is limited to 200,000 monthly users or $1 million in monthly revenue.

Why it matters

This model lowers the barrier for teams needing production-grade 3D geometry from a single image without paying for closed APIs. A 90% success rate, compared to 7–17% slicing rates in competitors, makes it significantly more practical for additive manufacturing. The increase in training assets from 500,000 to approximately 2 million and the sparse-voxel self-pruning decoder indicate a meaningful architectural step forward compared to previous open-source models like TRELLIS.2 and Hunyuan3D-2.1.

Highlights

• Three-pass sparse-voxel refinement at 512→1024→1536 resolutions; self-pruning decoder reconstructs surfaces at 8x finer scale
• 90% 3D printing slicing success rate (Pixal3D: 7%, TRELLIS.2: 17%); 83.5% preference rate in blind evaluations by 13 professionals across 562 samples
• Commercial license is free but capped at 200,000 monthly users or $1 million monthly revenue; output is geometry only, excluding color or texture

Critical take

The entire evidence base relies on a single X post by a tech commentator, lacking a peer-reviewed paper or reproducible benchmark. The 83.5% preference rate is statistically weak for the headline claim, as it is based solely on votes from 13 evaluators. The 'free' label also masks a strict commercial cap that impacts any product exceeding hundreds of thousands of users. Furthermore, the model’s exclusive focus on shape leaves texture pipelines as an unresolved dependency.