# On the Importance of Noise Scheduling for Diffusion Models

**Authors:** Ting Chen  
**arXiv:** [2301.10972](https://arxiv.org/abs/2301.10972)  
**Date:** 2023-01-26  
**Summary page:** https://www.summarizepaper.com/en/paper/2301.10972v1/

## Key points

- Noise scheduling is crucial for determining the effectiveness of diffusion models
- Optimal noise schedule varies depending on the specific task and image sizes
- Increasing image size requires a shift towards a noisier schedule due to increased redundancy in pixels
- Scaling input data by a factor of $b$ while keeping the noise schedule function fixed improves model performance
- Combining this approach with Recurrent Interface Network (RIN) leads to state-of-the-art pixel-based diffusion models for high-resolution images on ImageNet
- These models enable single-stage, end-to-end generation of diverse and high-fidelity images at 1024x1024 resolution without upsampling or cascades.

## Summary

In the study titled "On the Importance of Noise Scheduling for Diffusion Models," Ting Chen empirically investigates the impact of noise scheduling strategies on denoising diffusion generative models. The research reveals three key findings that shed light on the significance of noise scheduling in achieving optimal performance. Firstly, the study establishes that noise scheduling plays a crucial role in determining the effectiveness of diffusion models. The optimal noise schedule varies depending on the specific task, such as image sizes. This finding emphasizes the need for tailored noise scheduling strategies to maximize performance in different scenarios. Secondly, when increasing the size of images, it is observed that the optimal noise scheduling shifts towards a noisier schedule. This shift is attributed to increased redundancy in pixels as image size grows. Understanding this relationship between image size and noise scheduling allows for more accurate and effective denoising processes. Thirdly, the research proposes a simple yet effective strategy for improving diffusion models across various image sizes. By scaling the input data by a factor of $b$ while keeping the noise schedule function fixed (equivalent to shifting the logSNR by $\log b$), significant enhancements in model performance can be achieved. This approach proves successful when combined with Recurrent Interface Network (RIN), a recently proposed technique. The combination of this straightforward recipe and RIN leads to state-of-the-art pixel-based diffusion models for high-resolution images on ImageNet. Notably, these models enable single-stage, end-to-end generation of diverse and high-fidelity images at an unprecedented resolution of 1024$\times$1024 without requiring upsampling or cascades. Overall, this empirical study highlights the critical role played by noise scheduling strategies in denoising diffusion generative models. The findings emphasize tailoring noise schedules based on specific tasks and reveal insights into how image size impacts optimal noise scheduling choices.