Reproduce “Human Preference Score: Better Aligning Text-to-Image Models with Human Preference” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Xiaoshi Wu, Keqiang Sun, Feng Zhu, Rui Zhao, Hongsheng Li

Researchers from CUHK MMLab and SenseTime Research introduced the Human Preference Score (HPS) and a method to better align text-to-image models with human aesthetic preferences. By fine-tuning a CLIP model on a large dataset of human choices, their HPS-adapted Stable Diffusion model produced images that were preferred by users 74% of the time and significantly reduced common visual artifacts.

Key insights

  • Traditional metrics (IS, FID) and even raw CLIP scores show poor correlation with actual human aesthetic preferences for images generated by T2I models.
  • Fine-tuning a CLIP model on a dataset of explicit human preference choices for synthetic images creates a robust and generalizable score (HPS) that significantly aligns with human judgment.
  • Leveraging the HPS to explicitly train generative models to avoid "non-preferred" characteristics through negative prompting is an effective strategy for improving aesthetic alignment and mitigating common visual artifacts.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.