Reproduce “Personalizing Text-to-Image Generation to Individual Taste” 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
Anne-Sofie Maerten, Juliane Verwiebe, Shyamgopal Karthik, Ameya Prabhu, Johan Wagemans, Matthias Bethge
This research introduces PAM∃LA, a novel large-scale dataset and corresponding personalized predictor designed to align text-to-image generation with individual aesthetic preferences. The predictor achieves superior user-specific preference prediction and effectively steers generative models to produce images genuinely preferred by individual users over outputs from generic optimization methods.
Key insights
- Individual aesthetic preferences for AI-generated images can be accurately modeled by combining multimodal inputs, including user-specific embeddings and demographic data, surpassing generic preference prediction.
- Personalized reward models can effectively steer T2I generation to align with individual tastes, preserving realism and compositional nuance, unlike generic models that often degrade aesthetic quality.
- Consensus-driven reward models can inadvertently overfit to the subjective tastes of specific demographic groups, potentially reinforcing biases and producing less diverse or appealing outputs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.