Abstract

We present in this paper a novel post-training quantization (PTQ) method, dubbed 

AccuQuant, for diffusion models. We show analytically and empirically that quan- 

tization errors for diffusion models are accumulated over denoising steps in a 

sampling process. To alleviate the error accumulation problem, AccuQuant mini- 

mizes the discrepancies between outputs of a full-precision diffusion model and its 

quantized version within a couple of denoising steps. That is, it simulates multiple 

denoising steps of a diffusion sampling process explicitly for quantization, account- 

ing the accumulated errors over multiple denoising steps, which is in contrast to 

previous approaches to imitating a training process of diffusion models, namely, 

minimizing the discrepancies independently for each step. We also present an 

efficient implementation technique for AccuQuant, together with a novel objective, 

which reduces a memory complexity significantly from O(n) to O(1), where n 

is the number of denoising steps. We demonstrate the efficacy and efficiency of 

AccuQuant across various tasks and diffusion models on standard benchmarks.