>>8625915
>sho \(sho lwlw\)
>ningen mame
These were not present in the dataset at all, so that's about to be expected. Try using the lora extract on top of your favorite shitmix or even base noob (or you can even extract the difference and merge it to a model yourself) which should tamper with styles far less than the genning on the actual trained checkpoint while still keeping 1536x res.