Reproduce “The Algorithmic Gaze of Image Quality Assessment: An Audit and Trace Ethnography of the LAION-Aesthetics Predictor” 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
Jordan Taylor, William Agnew, Maarten Sap, Sarah E. Fox, Haiyi Zhu
The LAION-Aesthetics Predictor (LAP), a widely used model for curating generative AI datasets and evaluating image quality, embeds an "algorithmic gaze" reflecting imperial, realist, and male aesthetic biases. Empirical analysis revealed its strong preference for Western and Japanese art, realistic landscapes and portraits, and images depicting women, concurrently filtering out diverse cultural art forms and abstract styles, highlighting the critical need for pluralistic AI evaluation approaches.
Key insights
- The LAION-Aesthetics Predictor (LAP) encodes an "algorithmic gaze" that exhibits strong biases towards Western and Japanese art (imperial gaze), realistic imagery like landscapes and portraits (realist gaze), and images featuring women (male gaze).
- LAP's embedded biases stem from its development choices, including the equal weighting of training datasets with diverse collection contexts and inconsistent documentation, along with annotator demographics and the conflation of different aesthetic measurements in its training data.
- The study argues against a universal "gold standard" for aesthetic evaluation in AI, advocating instead for pluralistic and transparent metrics that acknowledge the subjectivity and cultural variability of aesthetic values.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.