Predicting human preference for generated image pairs using HPSv3 [P]

Hey! I'm looking for ways to predict human preference for a project I'm building. (imagebench.ai) I've tryed HPSv3 , https://github.com/MizzenAI/HPSv3 and made post about it here: https://imagebench.ai/blog/does-the-score-match-your-eye It looks ok, but have many limitation as you can see in my post.