Poisson Flow Generative Models

Authors: Yilun Xu, Ziming Liu, Max Tegmark, Tommi Jaakkola

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experimentally, PFGM achieves current state-of-the-art performance among the normalizing flow models on CIFAR-10, with an Inception score of 9.68 and a FID score of 2.35.
Researcher Affiliation Academia Massachusetts Institute of Technology {ylxu, zmliu, tegmark}@mit.edu; tommi@csail.mit.edu
Pseudocode Yes Algorithm 1: Learning the normalized Poisson Field
Open Source Code Yes The code is available at https: //github.com/Newbeeer/poisson_flow.
Open Datasets Yes For image generation tasks, we consider the CIFAR-10 [22], Celeb A 64 64 [38] and LSUN bedroom 256 256 [39].
Dataset Splits Yes We follow the training procedure in [33] and split the training data into 99% training and 1% validation sets for model selection.
Hardware Specification Yes All the experiments are run on a single NVIDIA A100 GPU.
Software Dependencies No The paper mentions 'Scipy library [37] with the RK45 [7] method' but does not provide specific version numbers for software dependencies.
Experiment Setup Yes We choose M = 291 (CIFAR-10 and Celeb A) 356 (LSUN bedroom), σ = 0.01 and = 0.03 for the perturbation Algorithm 2, and zmin = 1e 3, zmax = 40 (CIFAR-10) 60 (Celeb A 642) 100 (LSUN bedroom) for the backward ODE. We further clip the norms of initial samples into (0,3000) for CIFAR-10, (0,6000) for Celeb A 642 and (0,30000) for LSUN bedroom.