laion-synthetic-115m

This dataset is a version of laion-400m with filtering/replacement of noisy/inaccurate captions with captions generated via the BLIP model. Provided by salesforce in BLIP. Modified to be compatible with img2dataset tool.

Download captioned images

Note: you may want to change some of the keyword arguments depending on your specific needs.

# Download parquet file containing mapping of image-URL's -> captions
wget -c https://huggingface.co/datasets/laion/laion-synthetic-115m/resolve/main/laion_synthetic_filtered_large.parquet
pip install img2dataset
# Download as many URL's as possible into webdataset (tars of txt/jpg files). Can also specify `files` instead.
img2dataset laion_synthetic_filtered_large.parquet --image_size 320 --resize_mode 'keep_ratio' --caption_col 'caption'--input_format parquet --output_format webdataset

> Downloading starting now, check your bandwidth speed (with bwm-ng)your cpu (with htop), and your disk usage (with iotop)!