Rethinking Noise Sampling in Class-Imbalanced Diffusion Models.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Diffusion models are powerful for image generation but face challenges with long-tailed data distributions.
• Class imbalance in training data leads to a head-class accumulation effect in diffusion models' latent space.
• Existing noise sampling strategies can exacerbate bias towards dominant classes, degrading generation quality and diversity.

Purpose of the Study:

• To investigate the impact of noise sampling strategies on class imbalance in diffusion models.
• To propose a novel sampling strategy, Bias-aware Prior Adjusting (BPA), to mitigate class imbalance effects.
• To enhance the quality and diversity of generated images from imbalanced datasets.

Main Methods:

• Analyzing the head-class accumulation effect in diffusion model latent spaces.
• Developing the Bias-aware Prior Adjusting (BPA) sampling strategy.
• Implementing adaptive noise sampling distribution priors for each class during training.

Main Results:

• Consistent noise sampling amplifies bias towards head classes, negatively impacting generation.
• BPA effectively debiases diffusion models in class-imbalanced scenarios.
• Experiments show BPA significantly improves image diversity and quality compared to standard methods.

Conclusions:

• The proposed BPA strategy is effective in addressing class imbalance in diffusion-based image generation.
• BPA offers a practical solution for improving generative model performance on real-world, imbalanced datasets.
• This work contributes to more robust and equitable image generation capabilities.