From four views to a voxel sculpture.

The portrait now uses a learned 3D reconstruction. Hunyuan3D-2mv builds the head, face, neck, and shirt from four views of the generated turnaround. The model ran locally; no Meshy geometry is used.

How it is built

1. Crop the front, left profile, back, and right profile from the reference sheet and remove their backgrounds.
2. Run Tencent's Hunyuan3D-2mv shape model to infer a continuous surface. This supplies facial depth and the shape of the head, rather than carving a volume from silhouettes.
3. Convert that mesh into a solid grid of 174,057 voxels, 96 cubes per bust height, close to the cube size of the original artwork. Export the exposed cube faces with beveled edges.
4. Align the source images' hairline, eyes, nose, mouth, chin, and collar to the learned mesh. Sample colors from visible views, with smooth weighting between directions.
5. Remove the lighting painted into the references. For each view, fit its shading as a smooth function of surface direction and of how enclosed each cube is, then divide it out. Then even out painted highlights, stubble shadow and view seams within each material, and give each cube one clean color.
6. Render with real lights: a studio environment with a warm key from the upper left, a cool fill, and pale blue ambient light, matched to the original artwork. Shadows now stay put as the bust turns. The download contains the same voxel model.

The reference views are AI-generated, not calibrated photographs of a real sculpture. The side and rear details are inferred, and the texture alignment is approximate. This is not a recovery of the exact blocks in the original artwork.

Hunyuan3D source · Build report · Download the voxel mesh