Nepotistically Trained Generative-AI Models Collapse

TL;DR

This paper demonstrates that generative-AI models, when retrained on their own outputs, produce distorted images and struggle to recover, highlighting a collapse phenomenon caused by nepotistic training.

Contribution

It reveals a novel collapse behavior in generative-AI models retrained on their own outputs, showing persistent distortions and recovery challenges.

Findings

Retraining on generated images causes significant distortions.

Models struggle to recover even after retraining on real images.

Distortions extend beyond initial prompts and persist over time.

Abstract

Trained on massive amounts of human-generated content, AI-generated image synthesis is capable of reproducing semantically coherent images that match the visual appearance of its training data. We show that when retrained on even small amounts of their own creation, these generative-AI models produce highly distorted images. We also show that this distortion extends beyond the text prompts used in retraining, and that once affected, the models struggle to fully heal even after retraining on only real images.