Pseudorandom Streams within Diffusion Models Act as Learnable Inputs That Affect Generation Quality
2026-08-03 • Machine Learning
Machine Learning
AI summaryⓘ
The authors studied how the randomness used in diffusion models, which is actually generated by deterministic computer algorithms called pseudorandom sequences, can influence the models’ training and output. They showed that these sequences have patterns that a small neural network can predict, meaning the randomness isn’t completely random to the model. By testing with random data instead of real images, they found that different pseudorandom sequences change how well the model learns and creates images. This suggests the source of randomness acts like a structured input that affects the model’s behavior, not just a simple random noise source.
diffusion modelspseudorandom number generatorsfinite-precision hardwaremultilayer perceptronsequence predictabilityloss functionMNISTCIFAR-10power lawstochastic inputs
Authors
Shengzhi Deng, Chenqi Ye, Yanze Guo
Abstract
Diffusion models rely on stochastic inputs, yet on finite-precision hardware, the "randomness" they consume is realized as deterministic numerical orbits generated by pseudorandom rules. Accessible orbit structure can become a learnable input and affect both training and generation because the realized loss and its gradient depend on the concrete pseudorandom values consumed at each optimization step. A small multilayer perceptron predicts the next value of an orbit from its recent history, measuring general sequence predictability. A diffusion probe replaces real images with online random tensors while preserving the diffusion architecture and training objective, measuring whether the target system can exploit orbit structure. After controlling marginal statistics and screening out clear dynamical and finite-precision failures, the remaining orbits still produce markedly different diffusion losses and generation quality on MNIST and CIFAR-10. Both measures show strong rank correlations with macroscopic generation degradation, although their local rankings differ. After normalization by the IID baseline, the probe loss and the real-data diffusion loss approximately follow an empirical power law, with different exponents on the two datasets. These results suggest that a pseudorandom source is not only a distributional choice, but also a model-dependent structured input.