Reproduce “Personalize at Test Time: Learning User Preferences for Image Generation” 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

Jiamu Bai, Jiaming Hu, Yanhong Wu, Zellux Wang

PAT personalizes image generation from users’ past pairwise choices. It scores images across visual dimensions, fits user-specific weights, and steers a frozen diffusion model by optimizing its noise. On 50 held-out comparisons per user, PAT predicts 77.1% of choices for users with 31–60 training pairs, versus 69.3% for a population vector; generation alignment gains rely on GPT-5.5 judgments, not human evaluations.

Summary by alphaXiv.

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