---
dataset_info:
  features:
  - name: image
    dtype: image
  - name: conditioning_image
    dtype: image
  - name: text
    dtype: string
  splits:
  - name: train
    num_bytes: 11037324721.0
    num_examples: 50000
  download_size: 11046953921
  dataset_size: 11037324721.0
---
# Dataset Card for "hagrid50k"
This dataset is designed to train a ControlNet with human hands. It includes hand landmarks detected by MediaPipe(for more information refer to: https://developers.google.com/mediapipe/solutions/vision/hand_landmarker).
The source image data is from [HaGRID dataset](https://github.com/hukenovs/hagrid) and we use a modified version from Kaggle(https://www.kaggle.com/datasets/innominate817/hagrid-classification-512p) to build this dataset. We randomly select 50k data samples to build this dataset and the caption is generated using BLIP-2 model.