v1xl3t/ygo-card-art-gan

YGO GAN

A generative adversarial network trained from scratch to synthesize trading-card-style art. Built in spring of 2022.

Try it in your browser at the demo Space, which generates fresh images from this checkpoint.

Training

• Architecture. lightweight-gan (lucidrains), 128 x 128 output
• Data. 10,000+ unique crops of trading card art
• Compute. Trained for 48 hours on a Colab GPU, saving and evaluating every 500 steps
• This checkpoint. model_69.pt, the final checkpoint from run violetrun1, May 2022

The original training notebook is on Colab.

Generated samples

EMA outputs from this checkpoint, uncurated beyond picking a dozen.

Usage

pip install lightweight-gan

# place the checkpoint where the CLI expects it
mkdir -p models/violetrun1
# download model_69.pt from this repo into models/violetrun1/

lightweight_gan --name violetrun1 --load-from 69 --generate

Dataset note

The training images are card art owned by Konami. The dataset is not distributed here, only the trained weights and their outputs. This model is a research and portfolio project with no affiliation to Konami, and the weights are shared for non-commercial use.

Author

Vi Ellis. More work at vi-ellis.com.