Region-Based Semantic Factorization in GANs

Authors: Jiapeng Zhu, Yujun Shen, Yinghao Xu, Deli Zhao, Qifeng Chen

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

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct extensive experiments to evaluate our proposed method, mainly on two types of models, i.e., Style GAN2 (Karras et al., 2020b) and Big GAN (Brock et al., 2019). And the datasets we use are diverse, including FFHQ (Karras et al., 2019), LSUN bedroom, church, car (Yu et al., 2015), and Image Net (Deng et al., 2009). For metrics, we use Fr echet Inception Distance (FID) (Heusel et al., 2017), masked Mean Squared Error (MSE), and Identity loss (ID).
Researcher Affiliation Collaboration Jiapeng Zhu 1 Yujun Shen 2 Yinghao Xu 3 Deli Zhao 4 Qifeng Chen 1 1Department of CSE, The Hong Kong University of Science and Technology, Hong Kong, China. 2Byte Dance, Beijing, China 3Department of IE, The Chinese University of Hong Kong, Hong Kong, China. 4Ant Research, Hangzhou, China.
Pseudocode No The paper describes the steps of the method in paragraph form (e.g., "First, we need to compute the Jacobian... Second, obtaining Jf and Jb... Third, solving Equation (8) or Equation (11)...") but does not provide a formal pseudocode block or algorithm listing.
Open Source Code Yes Our source code can be found at here.
Open Datasets Yes And the datasets we use are diverse, including FFHQ (Karras et al., 2019), LSUN bedroom, church, car (Yu et al., 2015), and Image Net (Deng et al., 2009).
Dataset Splits No The paper mentions using pre-trained models and datasets but does not explicitly provide details about train/validation/test dataset splits used for their own evaluation.
Hardware Specification Yes All the experiments are conducted on a single RTX 2080 Ti GPU.
Software Dependencies No The paper mentions using pre-trained models from "Tensor Flow Hub" but does not specify software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes To handle such a case, we make a slight modification on JT b Jb to make it non-singular as JT b Jb JT b Jb + τtr(JT b Jb)I, (10) where I is the identity matrix and tr( ) denotes the trace. τ = 1e 3 is a small scaling factor.