PFGM++: Unlocking the Potential of Physics-Inspired Generative Models

Authors: Yilun Xu, Ziming Liu, Yonglong Tian, Shangyuan Tong, Max Tegmark, Tommi Jaakkola

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments show that models with finite D can be superior to previous stateof-the-art diffusion models on CIFAR-10/FFHQ 64 64 datasets/LSUN Churches 256 256, with median Ds. In class-conditional setting, D=2048 yields current state-of-the-art FID of 1.74 on CIFAR-10 without additional training.
Researcher Affiliation Academia 1Massachusetts Institute of Technology, MIT, Cambridge, MA, USA.
Pseudocode Yes Algorithm 1 EDM training Algorithm 2 PFGM++ training with hyperparameter transferred from EDM Algorithm 3 EDM sampling (Heun s 2nd order method) Algorithm 4 PFGM++ training with hyperparameter transferred from EDM Algorithm 5 DDPM training Algorithm 6 PFGM++ training with hyperparameter transferred from DDPM Algorithm 7 DDIM sampling Algorithm 8 PFGM++ sampling transferred from DDIM
Open Source Code Yes Code is available at https://github. com/Newbeeer/pfgmpp
Open Datasets Yes We consider the widely used benchmarks CIFAR-10 32 32 (Krizhevsky, 2009), FFHQ 64 64 (Karras et al., 2018) and LSUN Churches 256 256 (Yu et al., 2015) for image generation.
Dataset Splits No The paper mentions training, testing, and evaluation on datasets like CIFAR-10, FFHQ, and LSUN Churches. However, it does not explicitly state the specific training/validation/test splits (e.g., percentages or sample counts) used for these datasets in the provided text.
Hardware Specification Yes All the experiments are run on four NVIDIA A100 GPUs or eight NVIDIA V100 GPUs.
Software Dependencies No The paper mentions using specific architectures like 'NCSN++' and 'DDPM++', and the 'Adam optimizer', but it does not specify software dependencies with version numbers (e.g., PyTorch version, Python version, CUDA version).
Experiment Setup Yes For training, we utilize the improved NCSN++/DDPM++ architectures, preconditioning techniques and hyperparameters from the state-of-the-art diffusion model EDM (Karras et al., 2022). Specifically, we use the alignment method in section 4 to transfer their well-tuned hyper-parameters. For optimizers, following EDM, we adopt the Adam optimizer with a learning rate of 10e 4. We further incorporate the EMA schedule, learning rate warm-up, and data augmentations in EDM.