The purpose of this dataset was to distribute a high quality, free-to-use dataset containing samples that require no attribution and have an open license.
All of the images were captioned in a cluster containing:
38x 3090 24G
6x 4090 24G
8x A5000 24G
2x A100 80G
A couple volunteers running a 3090 or 4090.
The model was running in fp8 precision using 🤗Transformers and 🤗Accelerate for easy multi-GPU captioning.
The captioning was spread across 10 different… See the full description on the dataset page:
https://huggingface.co/datasets/bghira/photo-concept-bucket.
