Modulated Diffusion: Accelerating Generative Modeling with Modulated Quantization

Weizhi Gao, Zhichao Hou, Junqi Yin, Feiyi Wang, Linyu Peng, Xiaorui Liu

Abstract: Diffusion models have emerged as powerful generative models, but their high computation cost in iterative sampling remains a significant bottleneck. In this work, we present an in-depth and insightful study of state-of-the-art acceleration techniques for diffusion models, including caching and quantization, revealing their limitations in computation error and generation quality. To break these limits, this work introduces Modulated Diffusion (MoDiff), an innovative, rigorous, and principled framework that accelerates generative modeling through modulated quantization and error compensation. MoDiff not only inherents the advantages of existing caching and quantization methods but also serves as a general framework to accelerate all diffusion models. The advantages of MoDiff are supported by solid theoretical insight and analysis. In addition, extensive experiments on CIFAR-10 and LSUN demonstrate that MoDiff significant reduces activation quantization from 8 bits to 3 bits without performance degradation in post-training quantization (PTQ). Our code implementation is available at https://github.com/WeizhiGao/MoDiff.

Lay Summary:

Diffusion models are powerful generative frameworks capable of producing high-quality images. However, their generation process involves multiple iterative steps, leading to high computational costs. In this work, we investigate whether a unified framework can inherit the advantages of existing acceleration techniques to improve sampling efficiency.

We propose MoDiff, a general framework that not only inherits the advantages of existing caching and quantization techniques but also provides a unified solution for accelerating various diffusion models. Both theoretical analysis and experimental results demonstrate that MoDiff effectively enhances the performance of existing quantization methods.

MoDiff enables faster sampling, thereby facilitating the practical deployment of diffusion models and introducing a new direction for accelerating diffusion processes in research.

Link To Code: https://github.com/WeizhiGao/MoDiff

Primary Area: Deep Learning->Generative Models and Autoencoders

Keywords: Diffusion Models; Efficiency; Quantization; Caching Method

Submission Number: 13693