InstantHand

Pretrained weights for InstantHand: Feed-Forward 3D Gaussian Splatting for Animatable Hand Avatar Reconstruction from a Single Image (NeurIPS 2026).

InstantHand constructs an animatable 3D Gaussian hand avatar from a single image, given off-the-shelf MANO, camera, and mask estimates, without test-time optimization.

Files

Both checkpoints give identical results. Use either one.

File Size DINOv3 backbone Notes
checkpoints/instanthand_full.safetensors 1.4 GB included No separate DINOv3 download needed
checkpoints/instanthand_light.safetensors 0.2 GB not included Requires the DINOv3 ViT-L/16 weights (see the GitHub README)
DINOv3_LICENSE.md License of the DINOv3 backbone included in the full checkpoint

Usage

Download one checkpoint and save it where the code expects it:

pip install -U huggingface_hub
hf download coolho1129/InstantHand checkpoints/instanthand_full.safetensors --local-dir .
mkdir -p exps/neurips2026/stage2_best
mv checkpoints/instanthand_full.safetensors exps/neurips2026/stage2_best/model.safetensors

Then follow the evaluation instructions in the GitHub repository, for example:

python eval_interhand.py
python eval_hanco.py

The MANO hand model is not redistributed here. Register at https://mano.is.tue.mpg.de and follow the GitHub README.

Training data

The model was trained on InterHand2.6M, HO3Dv3, Goliath, Hand Appearance (HARP), and PALM, together with synthetic appearance data generated following Zhao et al. (CVPR 2025). The InterHand2.6M evaluation subject and all HanCo subjects are unseen during training. See Appendix B of the paper for details.

License

The InstantHand weights are released under CC BY-NC 4.0 for non-commercial research use, consistent with the licenses of the training datasets. The DINOv3 backbone included in the full checkpoint is distributed under the DINOv3 License (DINOv3_LICENSE.md). Use of MANO is subject to the MANO license.

Citation

@inproceedings{kim2026instanthand,
  title     = {{InstantHand}: Feed-Forward {3D} {Gaussian} Splatting for Animatable Hand Avatar Reconstruction from a Single Image},
  author    = {Kim, Chanho and Ali, Muhammad Salman and Gwak, Kyeonghwan and Shaheryar, Muhammad and Park, Incheol and On, Jeongwan and Baek, Seungryul},
  booktitle = {NeurIPS},
  year      = {2026}
}
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support