Called “consistent model reproducibility”

Reproducibility: different models produce same output

Two regimes

• Memorization regime: Memorizes training data but cannot produce new samples
• Generalization regime: Can produce new data samples

Provide theoretical justification for memorization regime

Study for unconditional diffusion models

An unconditional diffusion model generates new data samples without any specific conditions or guidance. It means the model learns the distribution of the training data and can generate new samples from this learned distribution, but it does not take any additional input or condition to direct the generation process.