Spatial-Contextual Discrepancy Information Compensation for GAN Inversion

Authors: Ziqiang Zhang, Yan Yan, Jing-Hao Xue, Hanzi Wang

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Both quantitative and qualitative experiments demonstrate that our proposed method achieves the excellent distortion-editability trade-off at a fast inference speed for both image inversion and editing tasks. Our code is available at https://github.com/ZzqLKED/SDIC. Experiments Experimental Settings Datesets. We evaluate our method on two domains: human faces and cars.
Researcher Affiliation Academia Ziqiang Zhang1, Yan Yan1*, Jing-Hao Xue2, Hanzi Wang1 1Xiamen University, China 2University College London, UK
Pseudocode No The paper describes its method and architecture in detail through text and diagrams (e.g., Figure 1) but does not include explicit pseudocode or algorithm blocks.
Open Source Code Yes Our code is available at https://github.com/ZzqLKED/SDIC.
Open Datasets Yes For the face domain, we adopt the widely-used FFHQ dataset (Karras, Laine, and Aila 2019) for training and the Celeb A-HQ dataset (Karras et al. 2017; Liu et al. 2015) for testing. For the car domain, we used the Stanford car dataset (Krause et al. 2013) for training and testing.
Dataset Splits No The paper mentions training and testing datasets but does not explicitly provide information on validation dataset splits or percentages.
Hardware Specification Yes Our model is trained 100,000 steps on the NVIDIA Ge Force RTX 3080 GPU.
Software Dependencies No The paper mentions using specific models and optimizers (e.g., 'ranger optimizer', 'Arc Face model', 'Res Net-50 model', 'Style GAN generator', 'e4e encoder') but does not provide specific version numbers for software dependencies or programming frameworks.
Experiment Setup Yes λLP IP S, λID, and λedit are empirically set to 0.8, 0.2, and 0.5, respectively. We use the ranger optimizer (Yong et al. 2020) with a learning rate of 0.001 and a batch size of 2. Our model is trained 100,000 steps on the NVIDIA Ge Force RTX 3080 GPU.