Reproduce “Vipera: Towards systematic auditing of generative text-to-image models at scale” 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
Yanwei Huang, Wesley Hanwen Deng, Sijia Xiao, Motahhare Eslami, Jason I. Hong, Adam Perer
Vipera introduces a human-AI collaborative system for systematically auditing generative text-to-image models at scale. It integrates interactive visual analytics, particularly hierarchical scene graphs, with AI-powered suggestions to help auditors uncover biases and issues by structuring exploration, comparing outputs across prompts, and fostering new auditing directions.
Key insights
- Hierarchical scene graphs provide an effective visual aid for structured, customizable auditing and sensemaking of large-scale T2I model outputs, helping auditors explore diverse criteria.
- Human-AI collaboration, where AI suggests new prompts and auditing criteria, can significantly broaden the scope of an audit and uncover overlooked issues.
- While supporting exploration and breadth, T2I auditing tools need to better balance broad exploration with capabilities for in-depth, context-aware analysis to address the vastness of the auditing space.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.