Q-Diffusion: Quantizing Diffusion Models

Diffusion models have achieved great success in image synthesis through iterative noise estimation using deep neural networks. However, the slow inference, high memory consumption, and computation intensity of the noise estimation model hinder the efficient adoption of diffusion models. Although post-training quantization (PTQ) is considered a go-to compression method for other tasks, it does not work out-of-the-box on diffusion models. We propose a novel PTQ method specifically tailored towards the unique multi-timestep pipeline and model architecture of the diffusion models, which compresses the noise estimation network to accelerate the generation process. We identify the key difficulty of diffusion model quantization as the changing output distributions of noise estimation networks over multiple time steps and the bimodal activation distribution of the shortcut layers within the noise estimation network. We tackle these challenges with time step-aware calibration and shortcut-splitting quantization in this work. Experimental results show that our proposed method is able to quantize full-precision unconditional diffusion models into 4-bit while maintaining comparable performance (small FID change of at most 2.34 compared to >100 for traditional PTQ) in a training-free manner. Our approach can also be applied to text-guided image generation, where we can run stable diffusion in 4-bit weights with high generation quality for the first time.

Traditional PTQ inference only goes through the quantized model once, while diffusion models quantization needs to address the accumulated quantization errors and varying activation distributions in the multi-time step inference. Addtionally, conventional PTQ approaches usually either use synthetic data that contains inconsistency with real inputs during the inference time or are not data-free. Our Q-Diffusion samples the calibration dataset with inputs uniformly from multiple time steps, which are an accurate reflection of the production data in a data-free manner.

The activation distributions of the concatenated deep feature channels (X₁) and shallow feature channels (X₂) in shortcuts differ significantly, resulting in a bimodal weight distribution that is challenging to quantize. We propose a "split" quantization scheme that performs quantization prior to concatenation, requiring negligible additional memory or compute.

@InProceedings{li2023qdiffusion,
  author={Li, Xiuyu and Liu, Yijiang and Lian, Long and Yang, Huanrui and Dong, Zhen and Kang, Daniel and Zhang, Shanghang and Keutzer, Kurt},
  title={Q-Diffusion: Quantizing Diffusion Models},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  month={October},
  year={2023},
  pages={17535-17545}
}