Pseudorandom Streams within Diffusion Models Act as Learnable Inputs That Affect Generation Quality
The paper argues that diffusion models can learn structure in the pseudorandom streams they are fed, and that this changes output quality.
On finite-precision hardware, the authors treat “random” inputs as deterministic orbits rather than pure IID noise. They test predictability with a small MLP and a diffusion probe that uses random tensors in place of real images. Even after filtering obvious numerical failures and matching marginal statistics, different pseudorandom orbits produced different losses and generation quality on MNIST and CIFAR-10. The authors report strong rank correlations between their measures and generation degradation, suggesting the random source functions as a model-dependent structured input.
ArXiv · AI/CL/LG's note
On finite-precision hardware, the authors treat “random” inputs as deterministic orbits rather than pure IID noise. They test predictability with a small MLP and a diffusion probe that uses random tensors in place of real images. Even after filtering obvious numerical failures and matching marginal statistics, different pseudorandom orbits produced different losses and generation quality on MNIST and CIFAR-10. The authors report strong rank correlations between their measures and generation degradation, suggesting the random source functions as a model-dependent structured input.
ArXiv · AI/CL/LG's note
score 4