Neural FFTs for Universal Texture Image Synthesis

Authors: Morteza Mardani, Guilin Liu, Aysegul Dundar, Shiqiu Liu, Andrew Tao, Bryan Catanzaro

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive evaluations confirm that our method achieves state-of-the-art performance both quantitatively and qualitatively. (Abstract)
Researcher Affiliation Industry Morteza Mardani , Guilin Liu , Aysegul Dundar, Shiqiu Liu, Andrew Tao, Bryan Catanzaro NVIDIA {mmardani,guilinl,adundar,edliu,atao,bcatanzaro}@nvidia.com
Pseudocode No The paper describes the network architecture and training process in textual descriptions and diagrams, but does not include structured pseudocode or algorithm blocks.
Open Source Code No The paper does not contain any explicit statement about releasing its source code, nor does it provide a link to a code repository.
Open Datasets Yes A large texture dataset with 55, 583 images from 15 different sources [8, 53, 9, 6, 7, 47, 1, 15, 45, 32] are collected. (Section 5)
Dataset Splits Yes The dataset is randomly split into a training set of 49, 583 images, a validation set of 1, 000 images, and a test set of 5, 000 images. (Section 5)
Hardware Specification Yes The model was trained on 4 DGX-1 stations with 32 total NVIDIA Tesla V100 GPUs and 320 CPUs using synchronized batch normalization layers [25]. (Section 5.1)
Software Dependencies No The paper mentions 'Pytorch interface with cu DNN' but does not specify version numbers for either software component.
Experiment Setup Yes We choose batch size of 8 per GPU, and the initial learning rate 10 5 that is halved every 200 epochs. Total of 800 epochs are used for convergence. We also set λvgg = 0.1, λstyle = 200, λadv = 0.1. (Section 5.1)