Cellpose-SAM WQuant — W8A8-QDQ

Ready-to-load W8A8-QDQ quantized variant of Cellpose-SAM cpsam_v2.

Load

import runpy
from huggingface_hub import hf_hub_download

loader = runpy.run_path(
    hf_hub_download("capicu-ai/cellpose-sam-wquant-w8a8-qdq", "runtime_loader.py")
)
model = loader["load_model"]()

Pass device="cuda" to load_model for CUDA.

Result

Weight size: 295.8 MiB
Compression versus FP32: 3.93x
Evaluated paired samples: 20
Complete bounded-panel criterion met: No
Reload verification normalized RMSE: 0.0

results.json and model_registry.json contain the compact machine-readable record.

Citation

If you use this model or code, please cite our paper:

@article{cruzromero2026retention,
  title={Retention-Constrained Post-Training Quantization of Cellpose-SAM for Stem Cell Microscopy},
  author={Cruz Romero, Sebasti{\'a}n A.},
  journal={arXiv preprint arXiv:2609.21038},
  note={Accepted at NeurIPS 2026 LXAI Workshop},
  year={2026}
}