Towards demystifying the creativity of diffusion models

July 15, 2026 | Source: Google Research Blog

Tags: diffusion models, generative AI, score function, ICLR 2026, Google Research, memorization, theoretical ML

Google Research's ICLR 2026 paper proves that diffusion models' ability to generate novel images is a mathematical consequence of neural networks learning a 'smoothed' score function — forcing interpolation between training examples along the data manifold rather than memorizing any single one.

Details

For years, diffusion models' capacity to generate genuinely new images — rather than recombining training data verbatim — was treated as a curious empirical fact without a rigorous explanation. Google Research scientists have now provided one in 'On the Interpolation Effect of Score Smoothing in Diffusion Models,' presented at ICLR 2026. The paper's core argument: neural network training naturally smooths the learned score function — the force field guiding noisy data toward clean outputs. That smoothing forces the generative process to interpolate between training data points along the underlying data manifold, producing outputs that fall between real examples rather than coinciding with any single one. The practical implication is that model creativity is not emergent randomness — it is a deterministic consequence of standard neural network optimization. When trained on cat photos, a diffusion model does not retrieve cats; it traverses a cat-shaped manifold, surfacing points no training photo occupies. This theoretical grounding has clear downstream relevance for memorization and copyright research: models that interpolate are structurally distinct from models that retrieve, with meaningfully different implications for intellectual property risk. Code accompanies the paper.