Ayakashi Triangle //ultimate(bad).
trigger words changed(see in model triggerwords if use non v2.0 ultimate)
Trigger words:
works +-fine:
suzukanade
matsurikazamakif
background
works usable:
shirogane
shiroganetrue
lucytsukioka
seigenkazamaki
ponosukeninokuru
small dataset, but usable:
suzumother
masuraosujimori
i don't know how to use it, it's completelly dead:
yayoitoba - drawn suzukanade. WHY?????? just use "yayoi"
souganinokuru - drawn ponosuke just use "souga" it's works
even if i found previously, this one is hard:
matsurikazamakim -
https://civarchive.com/models/42630?modelVersionId=49995 Use this One, works perfect
drawn female variant anyway, too small trigger word difference.
i found bad way to get close : <lora:AyakashiTriangleV2:0.7> (kim), 1boy, red hair
but it's bad anyway
https://civarchive.com/models/42630?modelVersionId=49995 Use this One, works perfect
What i learn from made that LoRa... DONT MADE BIG LoRa's with many characters.
Solo character if main(>50 images in dataset), or 3-4 if it's just third plane personages.
Why i upload it... because it's generating like 2 days(and manually dataset preparation +1.5days) and i don't want to do it again. I'ts useful for shirogane, lucy, ponosuke and maybe backgrounds(+ old LoRas characters). maybe i made 1 non useful personages as individual LoRas versions later.
strength 0.6-0.8, hiresfix or img2img for best results(it's works for every LoRa).
if you get bad results - try https://civarchive.com/models/4669/corneos-7th-heaven-mix this checkpoint and write in prompt position(like "front view, face"), because dataset contained not only front-face.
i focus on true anime style models for fanarts or fanmanga creation, if you want artstation ultra quality arts style this isn't your choise.
Train models isn't easy, crypto donations are welcome and you can reguest character in it:
Description
Most characters of Ayakashi Triangle anime.
Warning I changed trigger words in full version
Works fine with anime models like andromedaMix or grapefruit or https://civitai.com/models/4669/corneos-7th-heaven-mix
strength 0.6-0.8, hiresfix or img2img for best results(it's works for every LoRa).
Train models isn't easy(this model is 3days*12hours of dataset editing), crypto donations on some coffee are welcome, you can choise title for add to queue:
FAQ
Comments (3)
How do you put these two characters in the same image if they are of the same Lora?
Typing matsurikazamakif only gives us the old man
it works with my defaults(and on preview), if you [b]really[/b] need this model for some works i can try redo it to be better. (only because you write 2 messages with very big time difference, maybe you really need this model for some fanworks).
I don't really want to spend time on this model now. P.S. you can write matsurif, and it can works, maybe. (i do some mistakes on old LoRAs, and this one is one of non really good, but title unpopular, and you can get good results with some magic). For redo it i need to rethinks about tags to do it more distant and redo LoRA(few hours of GPU time)
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.