Assessing a Single Image in Reference-Guided Image Synthesis

Authors: Jiayi Guo, Chaoqun Du, Jiangshan Wang, Huijuan Huang, Pengfei Wan, Gao Huang753-761

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

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical results on various datasets demonstrate that RISA is highly consistent with human preference and transfers well across models.
Researcher Affiliation Collaboration Jiayi Guo1, Chaoqun Du1, Jiangshan Wang2, Huijuan Huang3, Pengfei Wan3, Gao Huang1,4* 1Department of Automation, BNRist, Tsinghua University, Beijing, China 2Beijing University of Posts and Telecommunications, Beijing, China 3Y-tech, Kuaishou Technology 4Beijing Academy of Artificial Intelligence, Beijing, China
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide concrete access to source code for the methodology described.
Open Datasets Yes We conduct our experiments on five datasets: Yosemite (Zhu et al. 2017), Celeb A-HQ (Karras et al. 2017), AFHQ (Choi et al. 2020), LSUN Church and Bedroom (Yu et al. 2015), all at the resolution of 256x256.
Dataset Splits No The paper describes how its training data is acquired ('Finally, for each dataset, 16k images with 16 different quality scores (k/16, k = 1, 2, ..., 15) is obtained as the training images') and describes 'testing samples' for human judgments, but it does not specify a conventional train/validation/test split for the RISA model itself.
Hardware Specification Yes The model is trained for 100 epochs using a single NVIDIA RTX 2080Ti GPU.
Software Dependencies No The paper states 'RISA is implemented in Py Torch (Paszke et al. 2019)' but does not provide specific version numbers for PyTorch or other software dependencies.
Experiment Setup Yes The batch size is set to 4 and the model is trained for 100 epochs using a single NVIDIA RTX 2080Ti GPU. We use the Adam (Kingma and Ba 2014) with β1 = 0 and β2 = 0.99. The weight decay and the learning rate are set to 10^-4. The weights of all modules are initialized using He initialization (He et al. 2015) and all bias are set to zero.