这个模型使用了超过500张高质量的原始图片进行训练。
这个模型拥有很好的画风模仿效果。训练图片全部来自该画师的原画集。
使用提示词“traditional media”可以获得马克笔手绘风格
Pony XL版本：https://civarchive.com/models/366668
推荐权重：0.6~1.0
推荐模型：Animagine XL V3
--- --- --- --- --- --- --- ---
This model used more than 500 high-quality original pictures for training.
This model has a good effect of restoring the artist's style. The training pictures are all from the original painting collection of the artist.
Use the prompt word "traditional media" to get the marker hand-drawn style.
Pony XL version：https://civarchive.com/models/366668
Recommendation Weight：0.6~1.0
Recommendation Model：Animagine XL V3
Description
FAQ
Comments (10)
女包拯挺好看doge
OK SO THATS how we get animagine to do the do - will have to do more studying now ;) sheesh this looks amazing <3
Thank you for your support, your model is great too~ ^_^
@flyx3 I have started to re-learn training on other models, what was your learning rate for so many steps? - Usually i do LESS steps at a higher LR
@duskfallcrew For large training sets (LORA model > 300 images), the learning rate is not particularly important. The loss rate is very common between 0.1 and 0.125, and 0.125 is not necessarily worse than 0.1.
@flyx3 I've just always used 5e4 on like 500+ images even on SDXL LOL, but I'm also very weird in that I just do what i know because other stuff hasn't always worked for me, so i'm curious because training wiht MORE steps has always made me panic. (Also that my colab said 9 hours for 5k steps eeesh)
@duskfallcrew This is the charm of model training, everything is unknown, but full of surprises ~ I wish you good luck!
What I use:
learning rate 1e-4
unet-lr 1e-4
text-encoder-lr 1e-5
@flyx3 Awesome, i've screenshot your settings comment and i'll see if i can calculate time/cost etc - it's clear with learning. rates that are LOWER you need longer :)