Pseudo Numerical Methods for Diffusion Models on Manifolds

Authors: Luping Liu, Yi Ren, Zhijie Lin, Zhou Zhao

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct unconditional image generation experiments on four datasets: Cifar10 (32 32) (Krizhevsky et al., 2009), Celeb A (64 64) (Liu et al., 2015), LSUN-church (256 256) and LSUN-bedroom (256 256) (Yu et al., 2016). To analyze the acceleration effect, we test Fenchel Inception Distance (FID (Heusel et al., 2018)) on different datasets under different steps and different numerical methods, including DDIMs, S-PNDMs, F-PNDMs and classical fourth-order numerical methods (FONs) (e.g., Runge-Kutta method and linear multi-step method).
Researcher Affiliation Academia Luping Liu, Yi Ren, Zhijie Lin & Zhou Zhao Zhejiang University {luping.liu,rayeren,linzhijie,zhaozhou}@zju.edu.cn
Pseudocode Yes Algorithm 1 DDIMs and Algorithm 2 PNDMs
Open Source Code Yes Our implementation is available at https://github.com/luping-liu/PNDM.
Open Datasets Yes We conduct unconditional image generation experiments on four datasets: Cifar10 (32 32) (Krizhevsky et al., 2009), Celeb A (64 64) (Liu et al., 2015), LSUN-church (256 256) and LSUN-bedroom (256 256) (Yu et al., 2016).
Dataset Splits No No explicit information on training/validation/test dataset splits is provided beyond using pre-trained models.
Hardware Specification Yes We use the 50-step, 512 batch size experiment on an RTX3090 to test the computational cost and the column time is the average computational cost per step in seconds.
Software Dependencies No No specific version numbers for software dependencies (e.g., libraries, frameworks) are provided.
Experiment Setup Yes The pre-trained models for Cifar10, LSUN-church and LSUN-bedroom are taken from Ho et al. (2020) and the pre-trained model for Celeb A is taken from Song et al. (2020a). In these models, the number of total steps N is 1000 and the variance schedule is linear variance schedule. We use the 50-step, 512 batch size experiment on an RTX3090 to test the computational cost and the column time is the average computational cost per step in seconds.