Google Research links diffusion-model novelty to score smoothing, not memorization

The ICLR 2026 work offers a mathematical account of why denoising models can produce new images and molecular candidates, but it does not yet supply benchmark evidence for production systems.

Google Research has published an explanation of research arguing that diffusion models generate novel outputs because neural-network training smooths their denoising guidance, rather than because generation randomly escapes memorized training examples.

The work, titled “On the Interpolation Effect of Score Smoothing in Diffusion Models,” was presented at ICLR 2026, Google Research said. It examines the score function, the learned component that guides a noisy sample through successive denoising steps toward a data-like result.

Diffusion training starts by adding noise to real examples, such as images, until they are unrecognizable. The model then learns to reverse that process. In an idealized system that learned only the individual training points perfectly, Google Research argues, denoising would return those same points. The system would behave as a retrieval mechanism and produce carbon copies rather than new samples.

The paper’s central claim is that neural-network training does not preserve that exact, point-by-point score function. It smooths the function. That smoothing creates an interpolation effect: denoising trajectories can move through regions between training examples and arrive at plausible new outputs.

Google Research describes the mechanism with a force-field analogy. During generation, data points move from noise toward meaningful structure, while the score function determines the direction of movement. An unsmoothed score field would direct points back to exact examples. A smoothed field can guide them toward previously unseen locations that still fit the learned distribution.

That distinction matters for teams evaluating diffusion systems in image generation, molecular discovery, and other tasks with complex local structure. It gives interpretability researchers a more precise question than whether a model is simply “creative”: how much score smoothing occurs, where it occurs in the denoising path, and how it affects the tradeoff between sample novelty and fidelity.

For product teams, the explanation should not be read as evidence that a diffusion model is safe from memorization or capable of reliable extrapolation. A model can generate outputs that differ from its training data while still retaining training examples, reproducing sensitive content, or failing outside the distribution represented in its data. Score smoothing describes a proposed mechanism for interpolation, not a blanket guarantee of originality, privacy, or scientific validity.

Uncertainty: Google Research’s summary does not provide the paper’s experimental benchmarks, model parameter counts, compute requirements, dataset details, or conference materials. It therefore remains unclear how the proposed account compares quantitatively with alternative explanations of diffusion-model generalization, or whether the effect predicts performance and memorization behavior in large production models.