Now accepting commissions! https://www.fiverr.com/s2/3403793c68
Always use Highres fix
Use ADetailer for Face
Images generated on RaenaMix (SD1.5) / AutismPonyXL (Pony)
General prompt (1.5):
<lora:NekoparaCoconut-v3-08:0.7>, ChopioCoconut, blonde hair, long hair, very long hair, hair between eyes, cat ears, animal ear fluff, cat girl, blue eyes, yellow eyes, heterochromia, slit pupils, tail, (looking at viewer:1.3),
mature female, large breasts,General prompt (Pony):
<lora:CoconutXL-v1-07:0.7>, ChopioCoconut, blonde hair, long hair, very long hair, hair between eyes, cat ears, animal ear fluff, cat girl, blue eyes, yellow eyes, heterochromia, slit pupils, looking at viewer,
large breasts, tail, dark-skinned female,Outfit 1 (Maid Outfit):
outfit_1, maid, lavender dress, frilled dress, lavender bow, maid headdress, neck bell, cleavage cutout, nametag, juliet sleeves, long sleeves, waist apron, frilled apron, garter straps, fishnets,Outfit 2 (Casual Outfit):
outfit_2, pink belt, neck bell, denim jeans, purple shirt, pink sweater, off-shoulder sweater, cropped sweater, buttons, shirt tucked in,Outfit 3 (Yukata):
outfit_3, purple yukata, floral print, green obi, wide sleeves,Outfit 4 (Santa Outfit):
outfit_4, santa costume, santa hat, mini hat, red hairband, neck bell, midriff, red shorts, short shorts, garter straps, fishnets, red footwear, red capelet,Description
FAQ
Comments (8)
Cool, how do you separate each concept, all the models I train are confused concepts, need a lot of prompts to play its original quality, and your model, only need a few prompts to perfectly restore the character. But the model seems to be a little bad at training images outside of the set
Usually people are told to remove/prune tags they want the model to always have but I do the exact opposite. I basically just tag every as I see it. I technically don't need to include the 'outfit_1' tags but I do this because it acts as a saftey net cover any tags I missed.
By doing this it allows you to prompt the outfits but also any other outfit you want making it flexible. Sometimes the outfits can leak over into non-official outfits. For example if an outfit has a stylised white shirt, if you ever prompt white shirt it'll look like the outfits so I try to use things that won't be used commonly. Things like 'shirt', 'skirt' are all completely removed
@turkey910 Cool My training method is very simple, just use enough data set, let the model learn a lot, summarize the rule. That is, no processing at all, tagging a lot of tags and letting the ai learn to generalize on its own, at the cost of being hard to fit.
For example, after each training of my model, I need to train a targeted detail to repair lora for fusion, which is time-consuming. The advantage is that the generalization is very high. As long as the concept can be derived from the large model, adding this lora will not affect it at all. Because my lora seems to be a painting concept for large models...
@Rnglg2 I also train using AnyLoRA which is super flexible
@turkey910 Is it easy to use? I will try to use it next time I train Maple's model =OwO=
@Rnglg2 Yeah. I use google colab so it's just one of the model drop downs on the options
@turkey910 I train locally but I have a big hard drive and I have close to 300 stable diffusion models available and every time I look for a model I get annoyed lol