Paper 2604.16044
Elucidating the SNR-t Bias of Diffusion Probabilistic Models
- Published
- Apr 2026
- Research lab
- Independent
- Citations
- 2
- GitHub
- 120 stars
01 In brief
Summary
This paper identifies a Signal-to-Noise Ratio-timestep (SNR-t) bias in Diffusion Probabilistic Models (DPMs), where during inference the SNR of denoised samples becomes misaligned with their timestep due to accumulated prediction and discretization errors.
The authors provide empirical evidence and theoretical proof showing that reverse-process samples consistently have lower SNR than forward-process samples at the same timestep, causing the network to overestimate noise predictions.
To mitigate this, they propose a training-free, plug-and-play method called Differential Correction in Wavelet domain (DCW).
DCW decomposes samples into frequency components using Discrete Wavelet Transform and applies dynamic differential correction to each component separately, leveraging the model's inherent denoising behavior (low-frequency first, then high-frequency).
Experiments on various models (IDDPM, ADM, DDIM, A-DPM, EA-DPM, EDM, PFGM++, DiT, FLUX, Qwen-Image) and datasets (CIFAR-10, CelebA, ImageNet, LSUN Bedroom) show significant FID improvements with negligible computational overhead.
The method also enhances existing bias-correction models like ADM-IP, ADM-ES, and DPM-FR.
02 From the paper
Abstract
Diffusion Probabilistic Models have demonstrated remarkable performance across a wide range of generative tasks. However, we have observed that these models often suffer from a Signal-to-Noise Ratio-timestep (SNR-t) bias. This bias refers to the misalignment between the SNR of the denoising sample and its corresponding timestep during the inference phase. Specifically, during training, the SNR of a sample is strictly coupled with its timestep. However, this correspondence is disrupted during inference, leading to error accumulation and impairing the generation quality. We provide comprehensive empirical evidence and theoretical analysis to substantiate this phenomenon and propose a simple yet effective differential correction method to mitigate the SNR-t bias. Recognizing that diffusion models typically reconstruct low-frequency components before focusing on high-frequency details during the reverse denoising process, we decompose samples into various frequency components and apply differential correction to each component individually. Extensive experiments show that our approach significantly improves the generation quality of various diffusion models (IDDPM, ADM, DDIM, A-DPM, EA-DPM, EDM, PFGM++, and FLUX) on datasets of various resolutions with negligible computational overhead. The code is at https://github.com/AMAP-ML/DCW.