Reproduce “Self-Consuming Generative Models Go MAD” 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
Sina Alemohammad, Josue Casco-Rodriguez, Lorenzo Luzi, Ahmed Imtiaz Humayun, Hossein Babaei, Daniel LeJeune, Ali Siahkoohi, Richard G. Baraniuk
The authors compare image-model loops using synthetic-only training, fixed real data, or fresh real data. They track precision as quality, recall as diversity, and FID across generations, including under sampling that favors polished outputs. Their experiments find that fixed real data can delay degradation, while fresh real data can prevent it in the studied settings, not universally.
Key insights
- Generative models can enter "autophagous loops" when trained on synthetic data created by previous generations, leading to a condition termed "Model Autophagy Disorder" (MAD).
- Sampling bias during synthetic data generation creates an inherent quality-diversity tradeoff: maintaining quality through cherry-picking accelerates the collapse of diversity.
- A continuous influx of fresh real data is crucial to prevent MAD, with a specific threshold ratio of fresh to synthetic data required for stable model performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.