A lora, for Krea 2, to make transformation-related images, trained for 7 days on a 1x8 5090 node on a dataset of about 20k images, includes speech bubble support this time.
Use the abliterated text encoder for best results.
supports basically everything transformation related, furry, tg, tf, inanimate, it's all there.
Best results adding as much detail to the prompt as possible.
Description
FAQ
Comments (11)
20k images, 7 days of training. i'll have to try this just for the superlatives-
While It's technically made to do tf stuff, it's more accurately a re-tune on everything that interests me and my personal aesthetic preferences; there's a ton of stuff in there.
Thanks for making this. If you ever decide yo put yourself through it again, could I propose including artist tags to help prompt a particular style? Thanks again.
Using lora probably the best way i can find so far (with mixed results), would be keen for an Anima version of this if it can be done
I did, i found it struggles to learn a lot of styles very effectively, for using another lora, try merging this into the base model first and loading a lora over that, might help
there is an abliterated TE for krea2 already?
I use qwen3VLInstruct4bHeretic_v10 and it's pretty good.
i know Krea 2 is better than anima. but anima is much cheaper and faster for me. is it possible to get this for anima?
If i had an extra thousand or so lying around to splash on compute xD
@randomusernameblablabla i'll just donate 1k buzz then
So how big of a dataset did you have and at what resolutions did you train? The models incredible the ability to learn transformation comics to this extent is wild. and i wanted to try something similar because my msin issue is it doesn't seem to do sfw transformation well at all. Krea2 is crazy i cant wait till we have finetunes of it.
Also did you handtag the dataset or use model captioning