Configuring Metadata for Training ControlNet with Custom Datasets

🧨 Diffusers
1

I want to train a ControlNet using my custom dataset, structured as follows:

train
├── conditioning_images
│   ├── 00001.png
│   └── 00002.png
└── images
    ├── 00001.jpg
    └── 00002.jpg

For the metadata.jsonl file, I understand it needs a file_name column linking to image files. However, when I provided a metadata.jsonl file like this:

{"file_name": "images/00001.jpg", "conditioning_image": "conditioning_images/00001.png", "text": "a man in a suit and tie speaking into a microphone"}
{"file_name": "images/00002.jpg", "conditioning_image": "conditioning_images/00002.png", "text": "a bald man in a suit and tie giving a speech"}

the load_dataset function raised an error:

ValueError: image at conditioning_images/00001.png doesn't have metadata in metadata.jsonl

I am wondering if there’s an issue with my metadata.jsonl file.