OMS-DPM: Optimizing the Model Schedule for Diffusion Probabilistic Models

Authors: Enshu Liu, Xuefei Ning, Zinan Lin, Huazhong Yang, Yu Wang

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

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate that OMS-DPM can find model schedules that improve generation quality and speed than prior state-of-the-art methods across CIFAR-10, Celeb A, Image Net, and LSUN datasets. When applied to the public checkpoints of the Stable Diffusion model, we are able to accelerate the sampling by 2 while maintaining the generation quality. Sec. 5: We experimentally validate OMS-DPM across a wide range of datasets, including CIFAR-10, Celeb A, Image Net-64 and LSUN-Church, and show that OMSDPM can achieve significantly better trade-offs on generation quality and speed than the baselines (Fig. 1).
Researcher Affiliation Collaboration 1Department of Electronic Engineering, Tsinghua University, Beijing, China 2Microsoft Research, Redmond, Washinton, U.S.A.
Pseudocode Yes Our complete search flow is shown at Algorithm 1. Time cost budget C should be given in advance.
Open Source Code Yes We have open-sourced our code at https://github.com/ jsttlgdkycy/OMS-DPM to allow the community to use OMS-DPM.
Open Datasets Yes We demonstrate that OMS-DPM can find model schedules that improve generation quality and speed than prior state-of-the-art methods across CIFAR-10, Celeb A, Image Net, and LSUN datasets.
Dataset Splits No The paper mentions 'training set' and 'validation set' in the context of splitting data for predictor training ('We split our generated schedule-FID dataset into two parts: training set and validation set.'), but it does not provide specific training/validation/test split percentages or sample counts for the main datasets (CIFAR-10, Celeb A, etc.) used to train the DPM models themselves.
Hardware Specification Yes The horizontal axis is the time cost of generating a batch of images evaluated on a single NVIDIA A100 GPU (10 NFEs is approximately equivalent to 1400ms latency).
Software Dependencies No The paper does not explicitly provide specific version numbers for software dependencies or libraries used in the experiments.
Experiment Setup Yes All models are trained with 128 batch size for 800k iterations, with a learning rate of 2 10 4. We use 0.1 as dropout ratio and 0.9999 as EMA rate.