A new study from Google AI Research attempts to quantify how diffusion models generate creative outputs. The team analyzed the latent space to determine if models truly innovate or simply interpolate between training samples. These findings provide a technical framework for developers to tune the balance between fidelity and novelty in generative AI.