This lora was trained on 60+ images from the artist Sakatsuki Yakumo. What drew me to this style was the colors and delicate shading. The lora itself listens to the prompt well enough. The hands and feet are decent as well. The backgrounds range from "decent" to "what the hell is happening"?
Version 3.0 was trained on Anima V1.0.
V2 is trained on Noob 1.0 V-Pred.
Version 1 was trained on AutismMix and I haven't tested on anything else.
To get the results you see in the images above, these are the settings I used.
Visit My Site For Models I Can't Post On CivitAI: https://qtgurl.com/
Description
FAQ
Comments (5)
I love the detail in this, it feels like the kind of detail we could get out of 1.5 but without all the weird issues. Would you be willing to post your training settings?
For training I use these settings: https://civitai.com/models/281404/lora-training-guide-anime-sdxl?modelVersionId=316795
In this case, I set the repeat to 15 with epoch of 6, saving at every 1 epoch. I use Kohya for training. The file link above doesn't set the caption extension to ".txt" as it should, just as a word of warning.
@Fish788 Thank you very much!! I am surprised at the repeats, I've seen a lot of people now only using repeat 1 and cranking up epochs. If you don't mind my asking, was your dataset split into subcategories or did you just test and find that gave a better training result?
@hasou It just depends on the results. If 15 repeats come out bad, I try 1 and crank up the epochs and vice versa. I haven't fount a singular training solution for all cases sadly. But a solid, consistent, dataset of around 50-100 images seem to be the ideal amount to work with. When I test a lora I look for prompt comprehension, style recognition, and consistent hands and feet similar to the base model it was trained on.
@Fish788 Thank you! That explanation actually made this 'click' a bit better for me, for some reason I was having a really hard time wrapping my head around appropriate repeat use. Gonna do some testing with this and see if I can make it work for me. Thank you again, I really appreciate it!!