A new paper from Google AI Research analyzes how diffusion models generate creative imagery. The team identifies specific mechanisms that drive visual novelty during the denoising process. This technical deep-dive clarifies why certain prompts trigger unexpected artistic leaps. Practitioners can now better tune sampling parameters to balance strict adherence to prompts with visual variety.