/vt/ - Virtual Youtubers

HLL-anon, I've got a request if you're around
I would have tried this myself, but since you already have the dataset, it'll be much faster for you

Like in picrel, Pony knows many chuubas, the character name was probably purged from training (or just renamed like the many artist tags)

A lora with this kohya config has a good chance of bridging the gap between the chuuba that the UNet knows and the character name absent in the TE
Just adjust the rank, epochs, etc to match the size of HLL dataset

I think the caption should be in this format: <chuuba>, tags describing everytihng in the image

Keep n tokens is set to 1 along with shuffle caption enabled, so that the <chuuba> captures the character itself

I tested this with a quick style lora for Pony, and it worked as this artist is very likely already known by Pony

kohya settings: https://files.catbox.moe/77lfbv.json
Style lora (Norza artist style): https://files.catbox.moe/6gbsqc.safetensors

>>71265558
nice diver La+
can you retry the same gen with "fisheye effect" or "fisheye_effect" in the negative?