AI Insight
This study demonstrates that AI image generators, specifically diffusion models, can learn patterns from the pseudorandom number sequences used during training, not just from the actual image data. The researchers found that different pseudorandom number sources produced measurably different output quality in generated images, with the models exploiting structural regularities in the random number streams themselves. Experiments on MNIST and CIFAR-10 datasets confirmed that the noise prediction component of diffusion models learns from both the training images and the specific pseudorandom sequences used to generate training noise.
Why it matters
This finding challenges the assumption that random number generators are interchangeable in machine learning systems and suggests that careful selection of pseudorandom sources could improve AI image generation quality. It also reveals a previously unrecognized factor that may contribute to reproducibility issues in training deep learning models.
Understand the Science
⚠️ Preprint – Noch nicht peer-reviewed
Dieser Artikel wurde noch nicht von unabhängigen Experten begutachtet. Die Ergebnisse sind vorläufig und sollten mit Vorsicht interpretiert werden.
Abstract: Digital learning systems consume concrete pseudorandom values rather than abstract random variables. These values enter the realized loss and its gradient during training. If a pseudorandom stream contains structure that is accessible to the model, this structure can therefore become part of what the learning system learns. We show that this effect can be strong enough to change generation quality in diffusion models. For diffusion noise prediction, there are two related routes by which the loss can be reduced. The model can learn regularities of the clean data and use them to infer the added noise. It can also exploit regularities of the noise source itself to recover the realized noise from the noisy input. Random-role ablation shows that the dominant source-dependent effect in our experiments is associated with the diffusion-noise roles. A diffusion probe that removes reusable real-image structure still shows clear source-dependent noise-prediction learning, and its source ordering closely matches that of real-data diffusion training. Different pseudorandom sources also produce large differences in generation quality on MNIST and CIFAR-10. Further experiments show that the same source structure can also be learned in an independent next-value prediction task, although this is not the learning mechanism used by the diffusion model. Changing the probe clean reference preserves much of the source ordering while changing the numerical response, producing clean-reference-dependent power-law relations. Overall, the results show that a pseudorandom stream is not only a source of stochastic variation: its concrete structure can act as a learnable input whose effect depends on the learning system.
Source: Pseudorandom Streams within Diffusion Models Act as Learnable Inputs That Affect Generation Quality