New Noise Schedule Improves Diffusion Model Performance on Imbalanced Data.

Researchers have identified a critical issue in score-based generative diffusion models when dealing with imbalanced datasets: low-frequency classes often suffer from poor generation quality and limited diversity. This occurs because low-density regions associated with these classes lead to inaccurate score estimations, and high-frequency classes tend to dominate the model's score space. To counteract this, the team proposes the Class-frequency Guided (CFRG) noise schedule. This novel approach strategically assigns larger-scale noises to low-frequency classes during the diffusion process. Experiments across various tasks, including image and text-to-image generation on imbalanced datasets like CIFAR-100-LT and ImageNet-LT, demonstrate substantial improvements, highlighting the importance of frequency statistics in noise schedule design.