README.md

metadata

dataset_info:
  features:
    - name: image
      dtype: image
    - name: conditioning_image
      dtype: image
    - name: text
      dtype: string
  splits:
    - name: train
      num_bytes: 11037324721
      num_examples: 50000
  download_size: 11046953921
  dataset_size: 11037324721

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 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.