<ul><li>Recent advances in generative models have made it hard to differentiate between real and synthetic data.</li><li>Self-consuming loops in training with synthetic data can lead to model collapse or instability.</li><li>Data curation based on user preferences can drive models to optimize those preferences, leading to converging distributions.</li><li>Study explores the impact of noisy and adversarially curated data on generative models, proposes attack algorithms for adversarial scenarios, and conducts experiments to demonstrate algorithm effectiveness.</li></ul>

Self-Consuming Generative Models with Adversarially Curated Data

Discover more