A FastGAN architecture for unconditional image creation. This project is based on the research outlined in the paper "Towards Faster and Stabilized GAN Training for High-fidelity Few-shot Image Synthesis". The original paper can be found at original paper.
The model was trained using the FastGAN implementation from the parent FastGAN Repository.
The dataset used in this project is available on the Hugging Face dataset hub at the following link: anime-sceneries. Users can download the dataset from the link and use it in their own projects. Alternatively, users can download the dataset using the following code:
from datasets import load_dataset
dataset = load_dataset("sulpha/anime-sceneries")The model was trained for 8 hours on a P100 GPU. (See parent repository for training details).
The model weights can be downloaded from here
Download the weights to the same root directory as eval.py
Example execution:
python eval.py --n_sample 2
where the n_samples is the number of images to be generated (>1 only). The script is modified to run on CPU only.