HPSv3: Towards Wide-Spectrum Human Preference Score

Summary

This paper talks about HPSv3, a new way to score how much people like images generated from text descriptions. It uses a large and diverse dataset of preferences and a special method to handle uncertainty in ranking the images, helping improve the quality of text-to-image generators.

What's the problem?

The problem is that existing methods to measure how good generated images are often don't capture the full range of human preferences or handle uncertain cases well, which limits how much they can improve the image quality.

What's the solution?

The solution was to build HPSv3, which trains on a wide range of human preferences and uses an uncertainty-aware ranking loss to better learn how to rank images even when preferences are unclear. This scoring system guides text-to-image models to refine their outputs step-by-step, producing better images over time.

Why it matters?

This matters because it helps create AI models that generate images people like more. By better understanding human preferences and improving image quality iteratively, HPSv3 can make text-to-image tools more useful and enjoyable for creative work and entertainment.