Enhancing Diffusion Models with Gaussianization Preprocessing

This paper introduces a novel approach to improve the performance of diffusion models by applying Gaussianization preprocessing to the training data. The core idea is to transform the data distribution to more closely resemble a Gaussian distribution, which simplifies the learning task for the model, especially in the early stages of reconstruction. This addresses the issue of slow sampling and degraded generation quality often observed in diffusion models, particularly with small network architectures. The method's applicability to a wide range of generative tasks is a significant advantage, potentially leading to more stable and efficient sampling processes. The paper's focus on improving early-stage reconstruction is particularly relevant, as it directly tackles a key bottleneck in diffusion model performance. Further empirical validation across diverse datasets and network architectures would strengthen the findings.

• •Gaussianization preprocessing can improve diffusion model performance.
• •The method addresses slow sampling and degraded generation quality.
• •The approach is applicable to a broad range of generative tasks.