InfinityGAN: Towards Infinite-Pixel Image Synthesis

Authors: Chieh Hubert Lin, Hsin-Ying Lee, Yen-Chi Cheng, Sergey Tulyakov, Ming-Hsuan Yang

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental evaluation validates that Infinity GAN generates images with superior realism compared to baselines and features parallelizable inference.
Researcher Affiliation Collaboration 1UC Merced 2Snap Inc. 3Carnegie Mellon University 4Yonsei University 5Google Research
Pseudocode Yes Figure 33: Implementation of spatial style fusion. We present (left) the original Style GAN2 forward function, and (right) a corresponding implementation for the spatial style fusion. We align the related code blocks on the left and right.
Open Source Code Yes All codes, datasets, and trained models are publicly available. Project page: https://hubert0527.github.io/infinityGAN/
Open Datasets Yes All codes, datasets, and trained models are publicly available. Project page: https://hubert0527.github.io/infinityGAN/
Dataset Splits Yes For image outpainting task, we split the data into 80%, 10%, 10% for training, validation, and test.
Hardware Specification Yes Note that training and inference (of any size) are performed on a single GTX TITAN X GPU. ... We perform all the experiments on a workstation with Intel Xeon CPU (E5-2650 2.20GHz) and 8 GTX 2080Ti GPUs.
Software Dependencies Yes We implement our framework with Pytorch 1.6, and execute in an environment with Nvidia driver version 440.44, cu DNN version 4.6.5, and Cuda version 10.2.89.
Experiment Setup Yes We use λar = 1, λdiv = 1, λR1 = 10, and λpath = 2 for all datasets. All models are trained with 101 101 patches cropped from 197 197 real images. ... We adopt the Adam (Kingma & Ba, 2015) optimizer with β1 = 0, β2 = 0.99 and a batch size 16 for 800,000 iterations.