Efficient calibration dataset sampling for post-training quantization of diffusion models

Diffusion models have emerged as a leading method for generating diverse and realistic samples, revolutionizing tasks like text-to-image generation, super-resolution, and inpainting. However, their slow sampling process hampers implementation. This work focuses on post-training quantization (PTQ) optimization for diffusion models. We propose an error-contribution-aware calibration dataset sampling scheme, effectively reducing calibration dataset size while preserving model precision. Experimental results demonstrate the effectiveness of our approach, achieving performance comparable to uniform sampling methods with a significant reduction in calibration dataset size.