Here is my first LoRa : Totally Spies , art style and characters !
Alex :
<lora:TotallySpies-18:1>, 1girl, short hair, black hair, small breasts, dark skin, dark-skinned female, smiling,
Clover :
<lora:TotallySpies-18:1>, 1girl, short hair, blonde hair, small breasts, purple eyes, smiling,
Sam :
<lora:TotallySpies-18:1>, 1girl, long hair, orange hair, small breasts, smiling,
Mandy :
<lora:TotallySpies-18:1>, 1girl, long hair, black hair, small breasts, mole under eye, makeup, smiling, frown,
Britney :
<lora:TotallySpies-18:1>, 1girl, long hair, blue hair, small breasts, purple eyes, smiling,
Jerry :
<lora:TotallySpies-18:1>, 1boy, short hair, grey hair, bald, thin mustache, suit,
For fun I generated them as kids, and they came out so cute and funny. You can also generate any character, and turn your daughter into a Totally Spy.
I trained it on 16 pictures of each character and 39 scenery.
It kinda works, but it's far from perfect. Not sure if I should have used more pictures, or more steps or something. Don't hesitate to share your advices. I also would like to see some of the good pictures you'll manage to make with it.
The model I used for the example pics is Anything 4.5. It also works well with some other anime models such as Ashmix, czeromix 2, dhClassicAnime, while with some other models it looks weird.
Thank the guy who made this easy : https://civarchive.com/models/22530/guide-make-your-own-loras-easy-and-free
by Dom83 - https://dom83.xyz/
Description
FAQ
Comments (10)
Which model did you use the LORA with for the example pics?
Anything 4.5
I'm glad you published because I had it on my backlog but I haven't gotten to build the dataset yet, so I won't now lol.
Regarding training, you might need more images, but I wouldn't do that before discarding other things. Mainly I think you should increase the dimension, at least to 64, although most styles go to 128, you are including characters and the style so people won't mind the extra MBs, and it should give better details overall. The other thing, what model did you use to train? For a flat cartoon style, what works better from experience and what I've seen is Anything v3.
Good work anyways, I'll try it when I get some time.
I trained it with the default model in the guide, it's called "animefull-final-pruned-fp16.safetensors".
Thanks for the advice, I did leave these settings as default, next time I'll set it to 128MB.
I add the dataset. It definitly could be improved with better screenshots.
Dude pls, CLover - blue eyes -_-
Oops, you're right, I did not pay attention to the colour of their eyes. '-_-
@Dom83 I see, Thanks a million for lora)
@Dantegonist How are your results so different than everyone else's? Is it your model? They're so solid.
@Jellaibit nono) I just use the minimum weight(0.4-0.6) and write a lot of different prompts for each character
Maybe you can add some of the other characters