StonyBrook-CVLab/PixCell-sample-data

PixCell — Sample Data

A small bundle of pre-extracted features for quickly testing PixCell sampling without preparing the full dataset.

It contains 32 patches (1024×1024) from 4 TCGA diagnostic whole-slide images across 4 cancer subtypes (BRCA, LUAD, COAD, PRAD), with the images, SD-3.5 VAE latents, and UNI2-h embeddings already extracted.

Layout

patches/
├── metadata/patch_names_all.hdf5          # index; key: tcga_diagnostic_1024
└── tcga_diagnostic/<subtype>/single_1024/<WSI>/<r_c>.jpeg
features/
└── tcga_diagnostic/<subtype>/single_1024/<WSI>/
    ├── <r_c>_sd3_vae.npy                   # (32, 128, 128) float16  (mean+std, 16 ch each)
    └── <r_c>_uni.npy                       # (16, 1536) float16      (4x4 UNI token grid)

Usage

Download the dataset and point data["root"] in your PixCell inference config.py (configs/pan_cancer/pixcell_256_inference.py / pixcell_1024_inference.py) at the extracted folder:

from huggingface_hub import snapshot_download
root = snapshot_download("StonyBrook-CVLab/PixCell-sample-data", repo_type="dataset")

Then run tools/sample_256.py / tools/sample_1024.py as described in the PixCell README.

Source & license

Images are derived from TCGA diagnostic slides, which are publicly available. Released under CC-BY-4.0.

Citation

@article{yellapragada2025pixcell,
  title={PixCell: A generative foundation model for digital histopathology images},
  author={Yellapragada, Srikar and Graikos, Alexandros and Li, Zilinghan and Triaridis, Kostas and Belagali, Varun and Kapse, Saarthak and Nandi, Tarak Nath and Madduri, Ravi K and Prasanna, Prateek and Kurc, Tahsin and others},
  journal={arXiv preprint arXiv:2506.05127},
  year={2025}
}