Model Card for AVA Image Clip Embeddings

The AVA image dataset is a collection of digital photos with ratings. It was used to create the visual scorer that evaluated the images in Laion 5B to create the the Laion-Aesthetics dataset

https://github.com/imfing/ava_downloader/

“AVA: A Large-Scale Database for Aesthetic Visual Analysis”.

Naila Murray, Luca Marchesotti, Florent Perronnin, CVPR 2012.

New aesthetics scorer: https://github.com/kenjiqq/aesthetics-scorer/

Original aesthetics scorer: https://github.com/christophschuhmann/improved-aesthetic-predictor/

They were processed with OpenClip BigG-14, L-14, and H-14 models.

• "laion/CLIP-ViT-bigG-14-laion2B-39B-b160k"

• "laion/CLIP-ViT-H-14-laion2B-s32B-b79K"

• "laion/CLIP-ViT-L-14-laion2B-s32B-b82K"

https://github.com/mlfoundations/open_clip

Not all images were processed!

Refer to the parquet for the succesfully processed images.

The parquet fields are

• "image_name", #same id as AVA csv
• "pooled_output"
• "projected_embedding"