Awesome

Towards Accurate Binarization of Diffusion Model

This implementation supports the paper "Towards Accurate Binarization of Diffusion Model". [PDF]

<img src="imgs/main.png" alt="main" style="zoom: 30%;" />

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 ours_util.py in the directory ./ldm/modules/diffusionmodules.
Place ddpm_ours.py in the directory ./ldm/models/diffusion
run bash train.sh

Main Results

Results under 4-bit activation quantization on LSUN-Bedrooms.

<img src="imgs/a4.png" alt="results" style="zoom: 33%;" />

Results for LDM-4 on LSUN-Bedrooms in unconditional generation by DDIM with 100 steps.

<img src="imgs/table.png" alt="table" style="zoom: 40%;" />

Visualization Results

Samples generated by the binarized DM baseline and BinaryDM under W1A4 bit-width.

<img src="imgs/samples.png" alt="samples" style="zoom: 23%;" />

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:

@misc{zheng2024accurate,
      title={Towards Accurate Binarization of Diffusion Model}, 
      author={Xingyu Zheng and Haotong Qin and Xudong Ma and Mingyuan Zhang and Haojie Hao and Jiakai Wang and Zixiang Zhao and Jinyang Guo and Xianglong Liu},
      year={2024},
      eprint={2404.05662},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}