hi3d-official/Hi3Dv1.0

Hi3D v1.0

English · 简体中文

High-resolution 3D geometry refinement built upon Sparc3D.

Live Demo · GitHub Repository · International Website · China Website

Planned open-source release: October 23, 2026.

This Hugging Face repository mirrors the official GitHub project preview. It contains project documentation, the system overview, and selected visual comparisons. No source code or model weights are included yet.

Overview

Hi3D v1.0 is the official Hi3D project for high-resolution 3D geometry refinement. Built upon the Sparc3D geometry pipeline, it uses a reference image together with a coarse mesh to refine object geometry through successive sparse reconstruction levels.

The project report describes training on approximately two million curated 3D assets and refinement up to an effective sparse extraction-grid resolution of 1536³. Hi3D v1.0 retains the Sparc3D representation and refinement architecture, with an emphasis on scaling training data and high-resolution geometry.

System overview

Image features condition sparse flow on coordinates derived from the coarse geometry. A self-pruning sparse VAE decoder expands the features around predicted surfaces, followed by mesh extraction. Between levels, the extracted mesh is revoxelized to provide the next sparse support.

Level	Nominal structural grid	Sparse extraction grid
1	64³	512³
2	128³	1024³
3	192³	1536³

These labels describe nominal sparse coordinate and surface-extraction resolutions. They do not represent a dense 1536³ latent tensor or a fixed output triangle count.

Visual comparisons

Local geometry

Red boxes mark the reference regions enlarged in each row. Shaded geometry and colored surface-normal renderings show local structures, including layered relief, architectural moldings, curved boundaries, scales, teeth, and concentric grooves. Hi3D v1.0 is highlighted in teal in the rightmost column. Click the figure to inspect the full-resolution image.

Multi-view comparisons

Front, back, and bottom views are shown for four objects. Results combine shaded geometry and colored surface-normal renderings; the color does not represent generated textures. Hi3D v1.0 is highlighted in teal.

More multi-view examples

Additional examples follow the same column order and rendering conventions as the preceding figure.

Scope and limitations

• Geometry output. Hi3D v1.0 focuses on geometry; it does not generate textures, materials, or vertex colors.
• Coarse-mesh dependency. An upstream coarse mesh supplies global structure. Severe structural errors in that initialization may persist after refinement.
• Occluded regions. Geometry hidden from the input image is a plausible completion rather than an exact reconstruction of unseen surfaces.
• Application-specific validation. Visual detail and nominal resolution do not by themselves guarantee mesh suitability for a particular downstream workflow.

Acknowledgments

Hi3D v1.0 builds upon Sparc3D: Sparse Representation and Construction for High-Resolution 3D Shapes Modeling. We acknowledge the original Sparc3D work and the research it builds upon.