metadata
language:
- en
license: mit
size_categories:
- 100M<n<1B
task_categories:
- unconditional-image-generation
This repository contains the dataset and resources related to the paper Modulated Diffusion: Accelerating Generative Modeling with Modulated Quantization.
The paper introduces Modulated Diffusion (MoDiff), an innovative, rigorous, and principled framework that accelerates generative modeling through modulated quantization and error compensation. MoDiff not only inherits the advantages of existing caching and quantization methods but also serves as a general framework to accelerate all diffusion models. Extensive experiments on CIFAR-10 and LSUN demonstrate that MoDiff significantly reduces activation quantization from 8 bits to 3 bits without performance degradation in post-training quantization (PTQ).