BiDM: Pushing the Limit of Quantization for Diffusion Models

This project is the official implementation of our “BiDM: Pushing the Limit of Quantization for Diffusion Models”. [PDF]

Requirements

Establish a virtual environment and install dependencies as referred to latent-diffusion.

Usage

• Replace the existing main.py in the LDM with our version of main.py.
• Place openaimodel_ours.py and util_ours.py in the directory ./ldm/modules/diffusionmodules.
• Place ddpm_ours.py and ddim_ours.py in the directory ./ldm/models/diffusion
• run bash train.sh

Main Results

• Results for LDM in unconditional generation by DDIM with 100 steps.

Visualization Results

• Samples generated by the binarized DM baseline and BiDM under W1A1 bit-width.

Comments

• Our codebase builds on latent-diffusion and stable-diffusion. Thanks for open-sourcing!

BibTeX

If you find BinaryDM is useful and helpful to your work, please kindly cite this paper:

@inproceedings{zhengbidm,
  title={BiDM: Pushing the Limit of Quantization for Diffusion Models},
  author={Zheng, Xingyu and Liu, Xianglong and Bian, Yichen and Ma, Xudong and Zhang, Yulun and Wang, Jiakai and Guo, Jinyang and Qin, Haotong},
  booktitle={The Thirty-eighth Annual Conference on Neural Information Processing Systems}
}