PQD: Post-Training Quantization for Efficient Diffusion Models

Diffusion models (DMs) have demonstrated remarkable achievements in synthesizing images of high fidelity and diversity. However, the extensive computational requirements and slow generative speed of diffusion models have limited their widespread adoption. In this paper, we propose a novel post-training quantization for diffusion models (PQD), which is a time-aware optimization framework for diffusion models based on post-training quantization. The proposed framework optimizes the inference process by conducting time-aware calibration. Experimental results show that our proposed method is able to directly quantize full-precision diffusion models into 8-bit or 4-bit models while maintaining comparable performance in a trainingfree manner, achieving a few FID change on ImageNet for unconditional image generation. Our approach demonstrates versatility and compatibility, and can also be applied to 512×512 text-guided image generation for the first time.