Elucidating the SNR-t Bias of Diffusion Probabilistic Models

TL;DR

This paper identifies a bias in diffusion probabilistic models related to SNR and timestep during inference, provides empirical and theoretical analysis, and proposes a correction method that improves generation quality with minimal overhead.

Contribution

It uncovers the SNR-t bias in diffusion models, analyzes its causes, and introduces a frequency-based differential correction technique to mitigate this bias.

Findings

The SNR-t bias causes error accumulation during inference.

The proposed correction method significantly improves model performance.

The approach works across various diffusion models and datasets.

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.