High-Resolution GAN Inversion for Degraded Images in Large Diverse Datasets

Authors: Yanbo Wang, Chuming Lin, Donghao Luo, Ying Tai, Zhizhong Zhang, Yuan Xie

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

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we present comprehensive experiments to evaluate our method. We experiment with image inpainting, image colorization and super-resolution. ... We conduct extensive experiments on restoration tasks with different datasets and GAN models. The experimental results demonstrate the superiority of our method against other inversion methods.
Researcher Affiliation Collaboration Yanbo Wang1*, Chuming Lin2, Donghao Luo2, Ying Tai2, Zhizhong Zhang1 , Yuan Xie1 1School of Computer Science and Technology, East China Normal University 2Tencent Youtu Lab
Pseudocode Yes Algorithm 1 Pseudocode of CRI
Open Source Code Yes Code is available at https://github.com/Booooooooooo/CRI.
Open Datasets Yes Image Net (Krizhevsky, Sutskever, and Hinton 2012) ... FFHQ (Karras, Laine, and Aila 2019) ... Celeb A-HQ (Karras et al. 2017) ... BSD100 (Martin et al. 2001) dataset and DIV2K (Agustsson and Timofte 2017) dataset.
Dataset Splits Yes We invert 1000 images from the validation set of Image Net, each from different classes, to quantitatively evaluate the methods.
Hardware Specification No No specific hardware details (e.g., GPU/CPU models, memory amounts) are mentioned in the paper.
Software Dependencies No The paper does not provide specific software dependencies with version numbers (e.g., library or solver names with version numbers like Python 3.8, CPLEX 12.4).
Experiment Setup Yes For Style GAN-XL with Image Net, we use 1000 iterations for the optimization stage and use 350 iterations for the finetune stage as in (Sauer, Schwarz, and Geiger 2022). For Style GAN2 with FFHQ, we use 500 iterations for the first stage and 20 iterations for the second stage as inversion with a face is simpler.