PTQD: Accurate Post-Training Quantization for Diffusion Models

Authors: Yefei He, Luping Liu, Jing Liu, Weijia Wu, Hong Zhou, Bohan Zhuang

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate that our method outperforms previous post-training quantized diffusion models in generating high-quality samples, with only a 0.06 increase in FID score compared to full-precision LDM-4 on Image Net 256 256, while saving 19.9 bit operations.
Researcher Affiliation Academia 1Zhejiang University, China 2ZIP Lab, Monash University, Australia
Pseudocode Yes Algorithm 1: Quantization noise correction.
Open Source Code Yes Code is available at https://github.com/ziplab/PTQD.
Open Datasets Yes We conduct image synthesis experiments using latent diffusion models (LDM) [45] on three standard benchmarks: Image Net[9], LSUN-Bedrooms, and LSUNChurches [60], each with a resolution of 256 256.
Dataset Splits No The paper mentions using datasets but does not explicitly provide details on training, validation, or test splits (e.g., percentages or sample counts for each split).
Hardware Specification Yes We have measured the latency of matrix multiplication and convolution operations in quantized and full-precision diffusion models using an RTX3090 GPU, as shown in Table 4.
Software Dependencies No The paper mentions "CUTLASS [24]" but does not provide specific version numbers for software dependencies or libraries used in their experiments.
Experiment Setup Yes All experimental configurations, including the number of steps, variance schedule (denoted by eta in the following), and classifier-free guidance scale, follow the official implementation [45]. ... LDM-4 (steps = 20 eta = 0.0 scale = 3.0) ... LDM-4 (steps = 250 eta = 1.0 scale = 1.5)