Pseudorandom Streams within Diffusion Models Act as Learnable Inputs That Affect Generation Quality

Imagine you're developing a generative model that creates images, like a digital artist. You want it to produce high-quality outputs, but the randomness it relies on can be tricky. Most models use standard random number generators, which can lead to predictable patterns that limit creativity and diversity in the generated images. This predictability can cause issues, such as generating similar outputs or failing to capture the complexity of real-world data, which is what's called overfitting. Essentially, the randomness isn't as random as it seems, and that can hurt the model's performance. To tackle this, the authors propose a method that treats the structure of these pseudorandom inputs as a learnable feature. They use a small neural network to predict the next value in a sequence of these inputs, allowing the model to adapt and potentially exploit the underlying patterns. By replacing real images with random tensors during training, they can assess how well the model can handle these structured inputs without losing its core functionality. Their experiments show that even slight variations in the input can lead to significant differences in the quality of the generated images, indicating that the randomness used is not just a technical detail but a crucial part of the model's design. This work shifts the focus from treating randomness as a mere distributional choice to recognizing it as a structured input that can be optimized, which could lead to more robust generative models in practice.

The approach introduces a novel way to leverage pseudorandom inputs in diffusion models.

The findings are supported by empirical results across multiple datasets, though some aspects may require further validation.