A new study from Google AI Research examines how diffusion models generate creative outputs. The team decomposes the sampling process to identify where novelty emerges versus where the model replicates training data. This analysis provides a framework for developers to tune the balance between fidelity and originality in generative tasks.