☯️ HY3D-Bench Demo

🧩 Part Dataset — structured decomposition

part/
  images/                # renderings
  water_tight_meshes/    # per-part meshes

hf download tencent/HY3D-Bench --repo-type dataset \
  --include "part/**" --local-dir ./hy3d-part

✨ Synthetic Dataset — AI long-tail (20 super / 130 / 1,252 sub)

Pipeline: LLM semantic expansion → Diffusion text-to-image → Image-to-3D textured mesh. For robustness, augmentation, zero-shot eval, simulation libraries.

synthetic/
  glb/   # textured meshes
  img/   # conditioning images

hf download tencent/HY3D-Bench --repo-type dataset \
  --include "synthetic/**" --local-dir ./hy3d-synth

3D preview pattern

Synthetic .glb files load directly in <model-viewer> / gr.Model3D. Example snippet (replace URL with a downloaded sample):

<model-viewer src="sample.glb" auto-rotate camera-controls
  style="width:100%;height:420px"></model-viewer>

🎁 Baseline — Hunyuan3D-Shape-v2-1 Small (0.8B)

Trained on Full-level data. Finetune of tencent/Hunyuan3D-2.1 shape DiT (3.3B → 0.8B). Shape-only (no Paint/PBR stage).

Files in tencent/HY3D-Bench

config.json  LICENSE-Model.txt  README.md
token2048/model_2048token.ckpt + .safetensor
token4096/model_4096token.ckpt + .safetensor

Python — inspect + shape inference (Hunyuan3D-2.1 pipeline + swap)

from huggingface_hub import HfApi
api = HfApi()
print(api.model_info("tencent/HY3D-Bench").siblings)

# Shape-only inference pattern (needs hy3dshape from Hunyuan3D-2.1):
# from hy3dshape.pipelines import Hunyuan3DDiTFlowMatchingPipeline
# pipe = Hunyuan3DDiTFlowMatchingPipeline.from_pretrained(
#     "tencent/HY3D-Bench", subfolder="token2048",  # or token4096
#     use_safetensors=False, device="cuda")
# mesh = pipe(image="input.png", num_inference_steps=30)[0]
# mesh.export("out.glb")

No config.yaml ships in HY3D-Bench — reuse the Hunyuan3D-2.1 DiT config as reference and match input_size to 2048/4096. Full ZeroGPU Gradio code is in gradio-upgrade/ in this repo.

🚀 Run inference (ZeroGPU path)

This Space is Static because the account can't currently host ZeroGPU (402: needs PRO or 30-day-old verified account). Two options:

1. Request a community grant for GPU hardware, or subscribe to PRO, then deploy gradio-upgrade/app.py to a Gradio + zero-a10g Space.
2. Duplicate the official Space tencent/Hunyuan3D-2 (RUNNING on ZeroGPU) and swap the checkpoint to tencent/HY3D-Bench.

# once eligible:
hf repos create YOURNAME/hy3d-bench-gradio --type space \
  --space-sdk gradio --flavor zero-a10g --public
hf upload YOURNAME/hy3d-bench-gradio gradio-upgrade \
  --repo-type space

See gradio-upgrade/DEPLOY.md in Files. Official references: Hunyuan3D-2, Hunyuan3D-2.1.

📎 BibTeX

@misc{hunyuan3d2026hy3dbenchgeneration3dassets,
  title={HY3D-Bench: Generation of 3D Assets},
  author={Team Hunyuan3D et al.},
  year={2026}, eprint={2602.03907}, archivePrefix={arXiv},
  primaryClass={cs.CV}, url={https://arxiv.org/abs/2602.03907},
}

Also cite P3-SAM (2509.06784), X-Part (2509.08643), Hunyuan3D-Omni (2509.21245), Hunyuan3D 2.1 (2506.15442), 2.0 (2501.12202), 1.0 (2411.02293). Full list on the model card.