Original title: Nepotistically Trained Generative-AI Models Collapse
Authors: Matyas Bohacek, Hany Farid
The article delves into AI’s ability to create coherent images based on massive human-generated data. Surprisingly, when these AI models are retrained using their own produced content, the resulting images become highly distorted. This distortion isn’t limited to the retraining text but impacts the entire image synthesis process. Even after retraining solely on authentic images, the models continue to struggle and fail to fully recover from this ‘poisoning’ effect caused by the self-generated data. The study highlights a concerning issue in AI image synthesis: models, once trained on their own output, face significant challenges in producing accurate and coherent images, persisting even after retraining with genuine data.
Original article: https://arxiv.org/abs/2311.12202