v3.0 released!
The dataset was improved by increasing the number of images and their diversity. The captions were also significantly improved.
Result? Much greater flexibility, less overfitting, and wider possibilities to use different styles or unusual compositions.
The mrsh_c1r1 trigger word should be enough. There's no need to add "Ciri"/"The Witcher 3 game" etc. to the prompt, as this can force the in-game screen style and reduce flexibility.
Prompt example:
mrsh_c1r1, woman, close-up beauty portrait, looking directly at the camera with a quiet confident expression, soft golden-hour sunlight through tall grass, warm luminous bloom, blurred meadow background, dreamy cinematic photography, 3:4 portraitIf you like this LoRA, be sure to give it a like, and if you’ve generated anything with it, share it with us in the gallery! Also feel free to leave suggestions for future character LORAs in the comments.
Description
changed in v2.0:
Fizgig → OneTrainer ADAMW
steps 1800 → 2250
result: a lot better details, less pixelated images
FAQ
Comments (9)
It looks great! Do you have any plans for yennefer too?
Thanks! Yup, Yennefer is next!
the lora only 200MB but got this OOM error ?
ERROR lora diffusion_model.tproj.1.weight Allocation on device
It's weird, but check with the new 2.0 version, which is trained with OneTrainer
@mrsh_loras yup it worked now with v2 ty
Thanks for this Lora. it's really helpful. 💕 Although it's a shame that it affects the overall style of the image so much, even when I reduce the strength.
Yeah, it was one of my first Loras for Krea2 ever and I wanted to preserve the characteristic elements of the Ciri's belt; maybe that's why I slightly overtrained the Lora, and I should release V3
@mrsh_loras That would be fantastic. Thank you for all your work
@convincedimpalawqcq465 done, v3 released and it should be better. Just read the new description and experiment with TextFusion Refusal-Reduction lora