Suggestions for use:
To obtain full strength for this LoRA, I'd advise using a style prompt like this one, which was used on the examples:
rendered in a vibrant digital gouache painting style with bold, expressive impasto brushstrokes varying from thick textured drags to feathery blends, multicolored gradient saturations bleeding dynamically across fabrics, hair, and skin surfaces, luminous subsurface glows on cheeks and lips, loose painterly edges with soft blurry contours and high-contrast rim lighting, simplified color-blocked backgrounds in complementary saturated washes, anime-inspired sharp-eyed expressive faces with glossy specular highlights and subtle reflective dots, high saturation pop art vibrancy, visible canvas texture and directional stroke marks for dynamic energy, ultra-detailed 8K masterpiece with cohesive brushwork and painterly cohesion.Description
FAQ
Comments (6)
Can you share what configuration you used? or if that is all the defaults, can you share tips on dataset and labeling?
Asking cause it looks amazing.
I do not do captioning on my LoRAs. I technically train with a trigger word (96yottea in this case) but it doesn't need to be included in the prompt. LoRA strength = 1.0
I used 95% original artwork from 96YOTTEA (please support the original artist, they're awesome) and some generations from civit.
Training config in AI-toolkit: rank 64, saved as BF16, generally 30-150 training images (100 in this case) and 30-100 epochs (4000-8000 steps total for all LoRAs). Learning rate = 0.0001 except for this LoRA where I went up to 0.00015 for a while. I also use the experimental differential guidance (set at "3"). Trained with 768 and 512 resolution buckets.
@xion1 Thanks so much for sharing. I really didn't know zimage could do this considering how distilled the model is.
which zimage base did you use?
@cobaltpixiv520 full bf16 for training and generation
@xion1 THANK YOU!