hunyuan3d-mini

Hunyuan3D-2mini — fast image-to-3D mesh

The 0.6B "mini" shape model from the Hunyuan3D-2 family, bundled with the delight and paint pipelines. ≈6–18 s per mesh on an RTX 5090, producing compact meshes under 1 MB.

What this repo is: the Hunyuan3D-2mini shape DiT, the 2.0-turbo DiT, and the delight / paint / paint-PBR pipelines — weights only, not a retrain. The samples below were generated with this build.

---

Samples

Single image in, mesh out. Source images were rendered with Z-Image Turbo; the meshes are Blender Workbench clay renders of the raw .glb output.

<table> <tr><td><img src="https://huggingface.co/ChrisColeTech/hunyuan3d-mini/resolve/main/samples/src-robot.png" width="330" alt="source photograph of a retro robot toy"></td><td><img src="https://huggingface.co/ChrisColeTech/hunyuan3d-mini/resolve/main/samples/mesh-robot.png" width="330" alt="generated mesh, front three-quarter view"></td></tr> <tr><td>_{**source** — `a cute chunky retro robot toy standing upright, rounded metal body, simple friendly face, studio product photograph on a plain white background`}</td><td>_{**mesh** — 0.72 MB `.glb`, generated in 18 s, seed 42}</td></tr> <tr><td><img src="https://huggingface.co/ChrisColeTech/hunyuan3d-mini/resolve/main/samples/src-teapot.png" width="330" alt="source photograph of a porcelain teapot"></td><td><img src="https://huggingface.co/ChrisColeTech/hunyuan3d-mini/resolve/main/samples/mesh-teapot.png" width="330" alt="generated mesh, front three-quarter view"></td></tr> <tr><td>_{**source** — `an ornate ceramic teapot with a curved spout and handle, glazed blue and white porcelain, studio product photograph`}</td><td>_{**mesh** — 0.70 MB `.glb`, generated in 6 s, seed 42}</td></tr> </table>

Textured (img2mesh_textured)

The same robot through the full chain — mini shape, then the bundled paint pipeline — instead of shape alone.

<table> <tr><td><img src="https://huggingface.co/ChrisColeTech/hunyuan3d-mini/resolve/main/samples/mesh-robot.png" width="330" alt="untextured shape output"></td><td><img src="https://huggingface.co/ChrisColeTech/hunyuan3d-mini/resolve/main/samples/mesh-robot-textured.png" width="330" alt="same mesh after the paint stage"></td></tr> <tr><td>_{**`img2mesh`** — 0.72 MB, **18 s**}</td><td>_{**`img2mesh_textured`** — 3.6 MB with a baked base-colour texture, **+85 s** for the paint stage}</td></tr> </table>

Turntable

Four views, 90° apart.

Recommended settings

Supported modes: img2mesh, txt2mesh, and the _textured variants (texturing is a separate paint stage — see gotchas)

Notes and gotchas

• This is the speed/size tier, not the detail tier. At 0.6B it returns meshes around 0.7 MB where a 1.5B model returns 15–38 MB for the same inputs — roughly 20–50× lighter. Silhouette, proportion and major features are captured; fine relief is smoothed away (in the robot sample the chest panel survives, the moulded nose does not). That trade is the point: it is the right model for drafts, previews and bulk generation, and the wrong one for hero assets.
• Sub-10 s generation makes it usable interactively — 6 s for a simple object, 18 s including model load on the first call.
• img2mesh output is untextured; texturing is a separate paint stage. The bundled delight/paint pipelines back the _textured job types and do produce correctly textured meshes. Paint cost scales with vertex count, and this model's small meshes make it the cheapest in the family to texture: 85 s here, against ≈14 min for a 3.3B mesh and ≈24 min for an 845k-face one. Most of that time is a pure-Python UV hole-filling loop, not GPU work.
• Give it a clean, centred subject on a plain background — background clutter costs geometry.

---

Files

---

Provenance

• Upstream base model: tencent/Hunyuan3D-2mini and the Hunyuan3D-2 paint/delight pipelines
• This build: redistributed as a split component tree. Weights are not retrained here.
• License: Tencent Hunyuan3D community license — refer to upstream for redistribution and commercial-use terms.