Same training dataset, network/conv dim/alpha were simply adjusted, halving the file size. It behaves much better with upscaling and has higher fidelity to the actual style.