The goal of these Google Colab notebooks is to capture the distribution of Steam banners and sample with a DCGAN.
- Run
download_steam_banners.ipynbto download Steam banners. - Run
GAN.ipynbto train one of the suggested GANs, preferably DCGAN, to generate Steam banners.
NB: If you need to pre-process data, e.g. to remove duplicates, have a look at remove_duplicates.ipynb.
The dataset consists of 27,974 Steam banners with RGB channels and resized from 460x215 to 28x28 resolution.
For faster computations, the dataset is aggregated and stored as a NumPy structure (.npy) in a folder called data/.
A DCGAN model was trained for 500,000 updates with a mini-batch size of 128, which is about 2,300 epochs.
After the training is done, the generator and discriminator networks are stored as .hd5 files in a folder called checkpoints/.
NB: A snapshot called 128x128.zip with 31,723 Steam banners, saved as .jpg files, with RGB channels and resized to 128x128 resolution,
is also provided. The data comes from another of my Github projects based on Steam API instead of SteamSpy API.
This dataset is not used here, due to constraints of Google Colab (available RAM, and computation time before the hosted
runtime is at risk to be killed), but it could be used in future projects such as StyleGAN for which image size is
expected to have square proportions and to be a power of 2.
Results obtained with different numbers of iterations are shown on the Wiki.
A grid of generated Steam banners after 500,000 iterations: 4k display
A grid of real Steam banners: 4k display
- Goodfellow, Ian, et al. "Generative adversarial nets." Advances in neural information processing systems. 2014.
- Radford, Alec, et al. "Unsupervised representation learning with deep convolutional generative adversarial networks." 2015.
- Keras implementations of Generative Adversarial Networks (GAN)
- Implementation of Deep Convolutional GAN
- Image Completion with Deep Learning in TensorFlow, in 2016.
- StyleGAN2:
- StyleGAN: