HPSv2-hf

This is a Huggingface CLIPModel flavor of the HPSv2 model, which is trained to predict human preferences over AI generated images.

I converted the model weights from the openclip format to huggingface CLIPModel. The two text and image embeddings were tested to be equal before and after conversion.

You can load the model the same as any huggingface clip model:

from transformers import CLIPProcessor, CLIPModel

model = CLIPModel.from_pretrained("adams-story/HPSv2-hf")
processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32") # uses the same exact vanilla clip processor

All credit goes to the original authors of HPSv2