Towards Diverse and Faithful One-shot Adaption of Generative Adversarial Networks

Authors: Yabo Zhang, mingshuai Yao, Yuxiang Wei, Zhilong Ji, Jinfeng Bai, Wangmeng Zuo

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments show that our method outperforms the state-of-the-arts both quantitatively and qualitatively, especially for the cases of large domain gaps. Moreover, our Di Fa can easily be extended to zero-shot generative domain adaption with appealing results.
Researcher Affiliation Collaboration 1Harbin Institute of Technology 2Tomorrow Advancing Life 3Peng Cheng Laboratory
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes Code is available at https: //github.com/YBYBZhang/Di Fa.
Open Datasets Yes For FFHQ adaption, the target images are collected from three datasets: (i) Artstation Artistic-face-HQ (AAHQ) [17], (ii) Met Faces [10], and (iii) face paintings by Amedeo Modigliani, Fernand Leger and Raphael [36]. For Cat adaption, we collect target images from the AFHQ-Wild validation dataset and divide them into Tiger, Fox, and Wolf datasets. We use AFHQ, Met Faces, and Artistic-Faces datasets during training, and their license are included in our repository https://github.com/YBYBZhang/Di Fa.
Dataset Splits No The paper mentions calculating metrics against "validation sets" for evaluation, but does not provide specific train/validation/test splits for the datasets used in their model's training process.
Hardware Specification Yes We finetune the generator for 300 400 iterations, which takes about 3 4 minutes on an RTX 2080Ti GPU.
Software Dependencies No The paper mentions using StyleGAN2, StyleGAN-ADA, e4e, pSp, CLIP, and ADAM optimizer, but does not provide specific version numbers for any of these software components or libraries.
Experiment Setup Yes For training, we use ADAM optimizer [13] with a learning rate 0.02 and set the batch size to 2. We finetune the generator for 300 400 iterations. Our overall training loss consists of three terms, i.e., the global-level adaption loss Lglobal, the attentive style loss Llocal and the selective cross-domain consistency loss Lscc. In our experiments, we use λlocal = 2 and λscc = max(0, niter n B Niter niter ).