Post
500
We uploaded huge human annotated preference dataset for image generation. Instead of just having people choose which model they preferer, we annotated an alignment score on a word by word basis for the prompt. rate the images on coherence, overall alignment and style preference. Those images that score badly were also given to annotators to highlight problem areas. Check it out!
Rapidata/text-2-image-Rich-Human-Feedback
We also wrote a blog post for those who want a bit more detail:
https://huggingface.co/blog/RapidataAI/beyond-image-preferences
We also wrote a blog post for those who want a bit more detail:
https://huggingface.co/blog/RapidataAI/beyond-image-preferences