I failed to make AI understand that i want one arm prostethic so made lora for it
IL V1 with both left/right in data: seems like IL also failed with this, so i will make separate right/left loras
Want to get acces to all current and future early acces models, request a model or just support me?
Comissions open on http://www.fiverr.com/s/gD2zEyX and https://ko-fi.com/renaissancegremlin
Here is my Patreon from 5$ supporter tier allow acces to ALL models that get posed with early acces: https://www.patreon.com/c/RenaissanceGremlin
My socials:
DISCORD SERVER
Twitter FB
My loras are also coming slowly to tensor art: https://tensor.art/u/836328663648595213
Description
FAQ
Comments (12)
fullmetal alchemist fans going crazy rn
what weight btw? your loras tend to LOVE changing the pic style
I feel IL loras generally tend to change the style quite much, that's the thing i have to give it to Pony that Pony is better in not messing up style by using concepts, while tryign to keep the style i usually don't go with strengh above 0.7
@Shrekman17 i'm sorry to disapoint i think lmao but this to me only happens consitently with your loras, you can on my page, pics with many loras, always same style, char loosk the same, with your lroas tho... not so much
@nobody32d So i have no idea what i'm doing wrong ;d
That might be because i'm doing all training on site since my current PC is not worth bothering with local set up, and training them on WAI NSFW instead of base IL since "base" IL for training on civit is for some reason still 0.1, and i feel using WAI fro training makes the loras better uality then 0.1 IL
@Shrekman17 sadly i can't help ya there since i don't train but i'd assume it's a train issue yes since it's only like VERY consistently with your loras specifically
@nobody32d The only idea i have is either training with checkpoint, or that i'm going for 1800+ steps and i see a lot of other people go for ~500 for IL loras, and with 1800+ that was my usual for Pony training, more steps for sure make it have influence on style but i feel loras made that way get concepts better, (and some of my new loras are made out of 100% gpt generations, those are for sure gonna have GPT style in them).
So i'm not sure what approach is better, and since i don't do a lot of generations outside of loras showcase i can't tell by myself
@Shrekman17 if ya need someone to make pics i can help on that part lol for now i can defenitely say it radically changes the picture's art style
@nobody32d That's my first test, https://civitai.com/models/1595714?modelVersionId=1806524 published epochs 5 and 15
@Shrekman17 Hopefully it works now man, rooting for ya
@nobody32d For now i gonna reduce my average steps to ~1500-1000, from 1900 to see if they gonna perform better/have less impact on the style, since that might be the best lead that i have that my models with 1900 steps are actually a bit overtrained and took more style then they should along ith concepts
@Shrekman17 also pls lemme know if ya manage to fix this, i really really like the right metal arm one