# Common Diffusion Noise Schedules and Sample Steps are Flawed

**Authors:** Shanchuan Lin, Bingchen Liu, Jiashi Li, Xiao Yang  
**arXiv:** [2305.08891](https://arxiv.org/abs/2305.08891)  
**Date:** 2023-05-15  
**Summary page:** https://www.summarizepaper.com/en/paper/2305.08891v4/

## Key points

- Existing diffusion noise schedules and sample steps have critical issues:
- Flawed designs do not enforce a zero signal-to-noise ratio (SNR) at the last timestep
- Some diffusion samplers do not start from the last timestep, causing discrepancies between training and inference stages
- Impact on model performance, especially in Stable Diffusion models generating images of varying brightness levels
- Rescale noise schedule for zero terminal SNR
- Train with v prediction
- Always start sampler from the last timestep
- Rescale classifier-free guidance to prevent over-exposure during sampling
- Aim of adjustments: Align diffusion process between training and inference stages for more accurate image samples reflecting original data distribution
- Validity of enforcing zero terminal SNR mathematically
- Common pitfalls in sampler implementations
- Visualizations demonstrate how different rescale factors affect image generation based on prompts
- Importance stressed on avoiding ϵ formulation in sampler implementations like DDPM sampling

## Summary

In their study "Common Diffusion Noise Schedules and Sample Steps are Flawed," researchers Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang address critical issues in existing diffusion noise schedules and sample steps. They reveal that these flawed designs do not enforce a zero signal-to-noise ratio (SNR) at the last timestep and some diffusion samplers do not start from the last timestep, leading to discrepancies between training and inference stages. This can greatly impact model performance, particularly in Stable Diffusion models which struggle with generating images of varying brightness levels. To remedy these problems, the researchers propose simple fixes such as rescaling the noise schedule to ensure a zero terminal SNR, training with v prediction, always starting the sampler from the last timestep, and rescaling classifier-free guidance to prevent over-exposure during sampling. These adjustments aim to align the diffusion process between training and inference stages, resulting in more accurate image samples that better reflect the original data distribution. The implementation section further demonstrates the validity of enforcing a zero terminal SNR mathematically while also highlighting common pitfalls in sampler implementations. Visualizations of sample steps showcase how different rescale factors can affect image generation based on specific prompts. The researchers stress the importance of avoiding ϵ formulation in sampler implementations like DDPM sampling. Overall, this study sheds light on crucial flaws in existing diffusion noise schedules and sample steps while offering practical solutions for improving model performance and generating diverse image samples across various brightness levels.