New Noise Schedule Improves Diffusion Models for Imbalanced Data

Jiequan Cui, Beier Zhu, Qingshan Xu, Xiaojuan Qi, Bei Yu, Hanwang Zhang· June 29, 2026 View original

▶ The 2-minute explainer

Summary

Researchers introduce the Class-frequency Guided (CFRG) noise schedule for diffusion models, which assigns larger-scale noises to low-frequency classes. This method significantly improves generation quality and diversity for imbalanced datasets by addressing issues of inaccurate score estimation and high-frequency class dominance.

Diffusion models, while powerful for generative tasks, often struggle with imbalanced datasets. Low-frequency classes tend to have sparse data regions, leading to inaccurate score estimations and suboptimal generation quality. Conversely, high-frequency classes can dominate the score space, further reducing the diversity and quality of samples from less common classes. A new study reveals a direct correlation between class frequency and the multi-scale noise schedule in diffusion models. Based on this insight, the proposed Class-frequency Guided (CFRG) noise schedule assigns larger-scale noises to low-frequency classes. This targeted approach helps alleviate the problem of low-density regions for underrepresented classes, ensuring more accurate score estimation. Experiments across various tasks, including image and text-to-image generation on imbalanced datasets like CIFAR-100-LT and ImageNet-LT, demonstrate substantial improvements over baseline methods, highlighting the critical role of frequency statistics in noise schedule design.

Why it matters

Professionals developing generative AI models, especially for real-world datasets that are often imbalanced, can use this technique to produce higher-quality and more diverse outputs, improving model fairness and utility.

How to implement this in your domain

  1. 1Analyze existing diffusion model training pipelines for performance on imbalanced datasets.
  2. 2Implement the Class-frequency Guided (CFRG) noise schedule in custom diffusion models.
  3. 3Experiment with different noise scaling strategies based on class frequency for specific datasets.
  4. 4Evaluate the impact of CFRG on generation quality, diversity, and fairness metrics for low-frequency classes.

Who benefits

MarketingEntertainmentHealthcareE-commerceDesign

Key takeaways

  • Diffusion models struggle with imbalanced datasets, leading to poor generation for rare classes.
  • A new CFRG noise schedule assigns larger noise to low-frequency classes.
  • This improves score estimation and prevents high-frequency class dominance.
  • Significant improvements in image and text-to-image generation on imbalanced data are observed.

Original post by Jiequan Cui, Beier Zhu, Qingshan Xu, Xiaojuan Qi, Bei Yu, Hanwang Zhang

"arXiv:2606.27696v1 Announce Type: new Abstract: In this paper, we are the first to examine the correlations between class frequency and the multi-scale noise schedule within diffusion models. For score-based generative models, low-density regions often lead to inaccurately estima…"

View on X

Originally posted by Jiequan Cui, Beier Zhu, Qingshan Xu, Xiaojuan Qi, Bei Yu, Hanwang Zhang on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses