Model created from memable anime facial expressions
Description
I had no idea how to call it better
FAQ
Comments (17)
need an illustrious version for each expression as well...man your Loras are too good, unfortunately, working with illustrious for some reason.
I started remaking my pony loras for IL, Guts is alredy done: https://civitai.com/models/1061368/berserk-scream-and-berserkingtime-face-or-shrekman-berserk-loras
In time i will re-make them all, with some of them the problem is that they were made when i had 0 idea how to actually make loras and they somehow turned out good, but with older ones i will for sure need to double check my data for IL version
@Shrekman17 that is great and waiting for them hope you will find a way as now you are a pro at it.
@krigeta At this point i'm just kinda know what i'm doing, instead of absolute 0 idea what i'm doing xD
@Shrekman17 hey if it is not too much to ask can you help me to train some loras?
@krigeta Sure, how can I help You?
@Shrekman17 I want to train high-quality Loras and have achieved some good results so far. However, creating a perfect one with consistently decent outcomes still feels out of reach. What tips do you have for tagging the images and optimizing training settings?
@krigeta Depends on what lora You want, for style in 90% cases is just make trriger word and then go with auto tag, the same goes for some character more complicated designs could take more effort.
Another general tip i can give for IL: if you use on site training you will see previews that are absolute garbage no matter how good is your lora and may think that lora is failed, but that's because training runs on IL 0.1, to get a proper check You need to use checkpoint, Hassaku/WAINSFW(it is also very good for SFW)/IL smooth are th most well rounded to check stuff on them.
The most important thing is your dataset and put the tags that You wan this lora to replicate from data.
My most common mistake was when i started was thinking that it is "the more the better" for Data, but optimal data is 30-50 images
@Shrekman17 Wow, amazing! I’m collecting images for the characters I’m working on. Can we talk in DMs? I won’t bother you too much—just simple questions about tagging and what to do if a LoRA fails. Is that okay?
@Shrekman17 I really want to learn how to create a good dataset, including what to do and what to avoid, but I also don’t want to spend all my time on it. I’m looking to find a sweet spot.
@krigeta sure, i'm at my last day at the job today before my days off so i will have way more time from tomorrow
@Shrekman17 that is great, really excited to work with you
The sanji heart eyes are killing me 🤣🤣🤣
Nice work!
yea