TRIAD

Rotation-Invariant Spherical Watermarking via Third-Order SO(3) Representation Coupling
ICML 2026

Pengzhen Chen, Yanwei Liu, Xiaoyan Gu, Antonios Argyriou, Wu Liu, Weiping Wang

This repository hosts the pretrained weights for TRIAD, a rotation-invariant spherical image watermarking method for panoramic ERP images.

Download

Using the Hugging Face CLI

pip install -U huggingface_hub
hf download JERCCC/TRIAD --local-dir checkpoints/triad-32bit

The downloaded directory should contain:

checkpoints/triad-32bit/
  model.pt
  config.json
  manifest.json

Using the TRIAD GitHub repository

If you have cloned the TRIAD code repository, you can also run:

pip install -e '.[hub]'
python scripts/download_checkpoint.py --repo-id JERCCC/TRIAD

The model expects 256 × 512 RGB ERP input and uses a 32-bit watermark.

License

The released model weights are provided under the MIT License.