Reproduce “Scaling Up Personalized Image Aesthetic Assessment via Task Vector Customization” 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
Jooyeol Yun, Jaegul Choo
A method for Personalized Image Aesthetic Assessment (PIAA) scales effectively by customizing task vectors derived from general image aesthetic (GIAA) and image quality (IQA) databases. The approach achieves superior generalization and higher Spearman's rank-order correlation coefficients (SROCC) on unseen datasets compared to meta-learning methods.
Key insights
- General image aesthetic (GIAA) and image quality (IQA) databases, despite not having explicit personalized annotations, inherently contain distinct aesthetic and quality biases that can be leveraged for personalization.
- Task vectors derived from different GIAA/IQA databases are largely orthogonal, indicating that each contributes unique knowledge, and combining them layer-wise allows for flexible and effective synthesis of personalized preferences.
- Personalization can be achieved in a highly parameter-efficient manner by learning only a small number of scalar coefficients for combining task vectors, enabling effective few-shot learning with minimal user data.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.