Original artist - https://www.pixiv.net/en/users/1500528
Version 7 is copying style the most, while v4 is still capable of doing something good, something even better than the v7, not sure which version is the best, so i'll just leave them both.
Version 2 is different, it contains ~33% scenery pictures in the dataset, so its very capable of doing some great scenery generations, instead of the older versions, but style is not that great, compared to them.
Base model for training - AbyssOrangeMix2_sfw, recommending to use it
You should also pick a prompt from the example pictures or the dataset, if you don't sure which will work