Google explains diffusion model creativity as a side effect of smoothing

Google Research published a post about On the Interpolation Effect of Score Smoothing in Diffusion Models, presented at ICLR 2026. The authors explain diffusion model creativity as a mathematical consequence of score smoothing.

Model creativity appears in the space between training examples

A diffusion model learns to denoise data: it turns random noise into a meaningful sample step by step. If it perfectly followed the score function implied by training examples, it would collapse into memorization and reproduce those examples.

Google argues that neural networks usually do not learn such perfectly sharp functions. Regularization such as weight decay, along with implicit effects of gradient-based training, produces a smoother version. That creates an interpolation zone where samples settle between known training points.

The work gives a cleaner language for generative AI originality

The useful part is that it replaces vague talk about creativity with a mechanism. In images or molecules, a diffusion model can find a new point on the data manifold because smoothing prevents collapse onto a training sample while still keeping the output in meaningful data space.

For teams using generative models, this is a practical frame for quality versus novelty. A model that only memorizes is a bad generator. A model that wanders too far from the manifold produces noise. The value sits in the controlled space between them.

A mathematical mechanism is not a license for anthropomorphism

The word creativity invites overstatement. Google itself describes the work as an initial effort and says more remains to be tested as distributions and architectures become more complex.

The technical meaning matters: this is not intent, taste or understanding. It is a property of training and data geometry that can lead to new plausible samples.

The next test is larger models and messier domains

The signal to watch is whether this mechanism holds for larger architectures, more realistic data distributions and domains beyond images. If it does, debates about memorization and originality in generative AI will get a firmer metric than whether an output looks impressive.

Lilith's verdict

Diffusion creativity here is not a muse. It is a particle that fails to land on an old photo and stops in the gap between two memories.