SignSparK checkpoints

Pretrained flow-matching checkpoints for SignSparK: Efficient Multilingual Sign Language Production via Sparse Keyframe Learning (ECCV 2026). One EMA checkpoint per body stream (unet_large, ~1.4B params each), trained on CSL-Daily, How2Sign and BOBSL.

• 💻 Code: https://github.com/JianHe0628/SignSparK
• 🗂️ Data: https://huggingface.co/datasets/LionelLow/SignSparK_data
• 📄 Paper: https://arxiv.org/abs/2603.10446 · 🌐 https://cogvis-cvssp.github.io/papers/signspark/

Files

hand/ema_0.9999_200000.pt
body/ema_0.9999_200000.pt
face/ema_0.9999_200000.pt

Usage

# fetch into $SIGNSPARK_CKPT_DIR
python tools/download_models.py --streams hand body face --dest ./checkpoints
export SIGNSPARK_CKPT_DIR=$(pwd)/checkpoints

# sample all three streams
python sample_all.py eval.note=CSLDaily eval.ode_stepnum=50 test_data='[CSL-Daily]'

See the code repo for training, sampling, and visualization.

License

Released for non-commercial research use, under the terms of the datasets the model was trained on (CSL-Daily, How2Sign, BOBSL).

Citation

@inproceedings{low2026signspark,
  title     = {SignSparK: Efficient Multilingual Sign Language Production via Sparse Keyframe Learning},
  author    = {Low, Jianhe and Symeonidis-Herzig, Alexandre and Ivashechkin, Maksym and Sincan, {\"O}zge Mercano{\u{g}}lu and Bowden, Richard},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026}
}