PetsGAN: Rethinking Priors for Single Image Generation
Authors: Zicheng Zhang, Yinglu Liu, Congying Han, Hailin Shi, Tiande Guo, Bowen Zhou3408-3416
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We construct abundant qualitative and quantitative experiments to show the superiority of our method on both generated image quality, diversity, and the training speed. |
| Researcher Affiliation | Collaboration | 1 University of Chinese Academy of Sciences 2 JD AI Research |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide any explicit statements about open-sourcing code, nor does it include links to a code repository. |
| Open Datasets | Yes | The first one is the Places50 dataset (Shaham, Dekel, and Michaeli 2019), which contains 50 natural landscape scene images, the second one is the LSUN50 dataset (Hinz et al. 2021), which contains 50 images with more complex scene images, and the third one is the Image Net50 dataset (Zhang, Han, and Guo 2021), which contains 50 images of objects. |
| Dataset Splits | No | The paper uses standard datasets (Places50, LSUN50, ImageNet50) and describes general training parameters, but it does not specify explicit training, validation, or test dataset splits (e.g., percentages or counts). |
| Hardware Specification | Yes | Our experiments are conducted on NVIDIA RTX 3090. |
| Software Dependencies | No | The paper mentions using Adam optimizer and VGG for initialization but does not provide specific version numbers for any software dependencies. |
| Experiment Setup | Yes | The input image I is resized to make the longer side no more than 256, then downsampled with scale factor 8 to obtain the low-resolution image c I. We adopt DGP (Pan et al. 2020) as the inversion method, in which disturbance is sampled from N(0, 0.5). The input noise of Pets GAN is sampled from N(0, 1). The window size s is set to 7. We use Adam optimizer for G, F, DG, D but with different learning rates. The training epoch of Pets GAN is 5000. The batch sizes for optimizing ρτ(δI, PF G) and ϕ(PG) are 1 and 32, respectively. The warm up strategy is used to train G and F. |