README.md

metadata

license: mit
dataset_info:
  features:
    - name: image
      dtype: image
    - name: label
      dtype:
        class_label:
          names:
            '0': gt
            '1': pose
            '2': st
    - name: caption
      dtype: string
    - name: gtimage
      dtype: image
    - name: stimage
      dtype: image
  splits:
    - name: train
      num_bytes: 1702123872.04
      num_examples: 15764
    - name: test
      num_bytes: 144819992.92
      num_examples: 1346
  download_size: 1762884199
  dataset_size: 1846943864.96
task_categories:
  - image-to-text
  - image-captioning
  - keypoint-estimation

Dataset Summary

The data is based on DeepFashion; turned into image pairs of the same person in same garment with different poses.

This won't preserve the person/garment at all but just want to process the data first and see what kind of controlnet it can train as an exercise for training a controlnet.

The controlnet_aux's openpose detector sometimes return black images for occluded human images so there won't be a lot of valid image pairs.