Temporal Dynamic Quantization for Diffusion Models

Authors: Junhyuk So, Jungwon Lee, Daehyun Ahn, Hyungjun Kim, Eunhyeok Park

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our extensive experiments demonstrate substantial improvements in output quality with the quantized diffusion model across various datasets.
Researcher Affiliation Collaboration 1 Department of Computer Science and Engineering, POSTECH 2 Graduate School of Artificial and Intelligence, POSTECH 3 Squeeze Bits. Inc
Pseudocode No The paper describes methods and equations, such as the quantization function (Eq. 5), but does not contain any explicitly labeled 'Pseudocode' or 'Algorithm' blocks.
Open Source Code Yes The source code is available at https: //github.com/ECo Lab-POSTECH/TDQ_Neur IPS2023.
Open Datasets Yes For the DDIM experiments, we utilized the CIFAR-10 dataset [45] (32x32), while for the LDM experiments, we employed the LSUN Churches dataset [46] (256x256).
Dataset Splits No The paper mentions selecting checkpoints with the 'lowest validation loss' in the context of QAT, which implies the use of a validation set. However, it does not provide specific details on the size or methodology of the training/validation/test dataset splits, beyond a 'calibration set of 5120 samples for PTQ'.
Hardware Specification Yes All of experiments were conducted on the high performance servers having 4x A100 GPUs and 8x RTX3090 GPUs with Py Torch [49] 2.0 framework.
Software Dependencies Yes All of experiments were conducted on the high performance servers having 4x A100 GPUs and 8x RTX3090 GPUs with Py Torch [49] 2.0 framework.
Experiment Setup Yes The models were trained for 200K iterations on CIFAR-10 and LSUN-churches, with respective batch sizes of 128 and 32. The learning rate schedule was consistent with the full precision model.