Trained from Necomi's work, based on PonyXL v6

V02, re-did dataset, trained with smaller dim