> However, the trade-off with our approach is that it requires a more complex loss that operates given only generated samples.