How Diffusion Models Achieve Generalization: Insights from a CVPR 2026 Tutorial

Generalization paradox of diffusion models

Diffusion models achieve state‑of‑the‑art results in image, video and audio generation. The training objective has a unique closed‑form solution that depends only on the training data, which would allow perfect memorization.

Mechanisms that enable generalization

Score smoothing : smoothing the estimated score function improves robustness beyond the training distribution.

Neural‑architecture inductive bias : design choices in the network impose priors that favor functions that extrapolate.

Training dynamics : the trajectory of optimization influences the model’s ability to extrapolate to unseen data.

Data geometry : the structure of the data manifold shapes the diffusion process and affects generalization.

All mechanisms are interpreted through the lens of optimal denoising, providing a unified explanation of how diffusion models can generate novel content despite the memorization‑appearing objective.

Practical Jupyter notebooks illustrating the concepts are available at https://analytic-diffusion.github.io/.

Code example

来源：专知
本文
约1000字
，建议阅读
5
分钟
深度扩散模型究竟是如何实现泛化的？