Lianima for Krea2
So, i dropped the dataset for Lianima as-is in the on-site trainer and did 10 epochs of training on Krea2 just to experiment. In the end, i am quite happy with the result which will allow me to test some Krea2 LoRA on top of my own style! 💕
Since i can't run it locally yet, i need to make it available to test it on-site (i don't have a membership anymore, i can't test from the trainer). (That's no longer the case, yeah!)
NB: since it is an ANIMA dataset, it's tag based, sorry for the NLP lovers, still not my forte 😅
NB2: The LoRA has a strong bias toward female character and Liadora base outfit (naked shoulder dress, black clothes).
Usage (personal preference):
Steps: 10
CFG: 1.2
Sampler: ER_SDE
Scheduler: Beta (or smoothstep for sd.cpp)
All showcases pictures (except the first few tests) are generated using stable-diffusion.cpp (which now support Krea2, yeah! Local gen is back!)
TE: https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct-GGUF/blob/main/Qwen3VL-4B-Instruct-Q8_0.gguf
Turbo: https://huggingface.co/realrebelai/KREA-2_GGUFs/blob/main/TURBO/Krea-2-Turbo-Q5_K_S.gguf
The model should also work well on Raw, in this case, go for CFG 3.5 and 30 steps from my first tests.
Version difference
Vanilla is an average of epoch 6-7-8-9 of the pure training result. It suffers from censorship obviously even if some NSFW pictures were present and the "classic" LoRA filter did not help (i tried several and even fiddled a bit with numbers). But this LoRA by @alcaitiff did! 💜
Uncensored is an ugly merge of the Mystic XXX LoRA at strength 0.8, both using standard layers addition and SVD. Since neither was exactly what i wanted, i re-did a merge between those. I did this merge for personal usage as i would rather only load one LoRA and use base Turbo. You can either use Vanilla on an uncensored model or use the source Lora which is great! 😍 (I may do a model merge and then a quantization to have LoRA-included version locally, not quite sure it would be useful to provide it here).
Thanks for reading! 💜
Description
Posted on Civitai, upload found elsewhere are FORBIDDEN and done without the author permission.
FAQ
Comments (2)
Did you try this? Apparently it's removing the filter in Krea 2 allowing for realistic non sfw generations. However it's not working for anime, seems like Krea 2 was not trained on non sfw anime, so I don't know if it's working with your semi-realistic style.
That's one of the things i tried among several test indeed, but with other knobs. I'll take a look at this one thanks!