Boosting Diffusion Models with an Adaptive Momentum Sampler

Authors: Xiyu Wang, Anh-Dung Dinh, Daochang Liu, Chang Xu

IJCAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experimental results on multiple benchmarks demonstrate that our proposed reverse sampler yields remarkable improvements over different baselines.
Researcher Affiliation Academia Xiyu Wang, Anh-Dung Dinh, Daochang Liu, Chang Xu School of Computer Science, Faculty of Engineering, The University of Sydney, Australia xiyuwang.usyd@gmail.com, adin6536@uni.sydney.edu.au, daochang.liu@sydney.edu.au, c.xu@sydney.edu.au
Pseudocode Yes Algorithm 1 illustrates the adaptive momentum sampler process.
Open Source Code Yes The code is publicly available at github.com/Shiny Gua/DPMs-with-Adam
Open Datasets Yes We utilize CIFAR10 [Krizhevsky et al., 2009] (32 × 32), Celeb A [Liu et al., 2018] (64 × 64), Image Net [Deng et al., 2009] (64 × 64), LSUN [Yu et al., 2015] (256 × 256) and Celeb A-HQ [Karras et al., 2017] (256 × 256) in experiments.
Dataset Splits No The paper states the datasets used and sampling steps, but does not specify train/validation/test splits, absolute sample counts for splits, or cross-validation setup for reproducing the data partitioning.
Hardware Specification Yes Our experiments run on one node with 8 NVIDIA A100 GPUs.
Software Dependencies No The paper discusses various models and frameworks (DDIM, DDPM, Analytic-DPM, LDM) and states that pre-trained models are used, but it does not provide specific version numbers for software dependencies such as programming languages, libraries, or frameworks.
Experiment Setup Yes The sampling steps are set to 4000 on Image Net and 1000 on other datasets. For the Eq. 5, we select three η values... All the hyperparameters for the sampling process are presented in Appendix. Also, Table 4 investigates the effects of b and c hyperparameters.