support me on kofi : https://ko-fi.com/n0utis
you like my model? follow me!
lora train compatible anime/cartoon/realistic
Without prompt:>1.0 with his outfit, variant using lora's prompt suggestions
Without prompt: > 0.5 without vanilla outfit, compatible with other lora and considered without outfit, so all you have to do is give him an outfit and a pose like any lora.
Have fun, give me suggestions for lora to create in the comments :)
Description
Comments (6)
Please take a look at the character, she's not the same, also you don't need to make the lora as large as 200 Mb, 50 Mb is enough, also you should have more trigger words for the character and her outfit, especially the eye and hair color
wd 1.4 puts the eyes and the outfit, with tag verification. what you see is not the result you will have, everything will depend on the strength of the lora and the basemodel used. depending on the basemodel you will certainly need to use a low weight (0.5) or high (1.2-1.5) for the characters, the dimensions must be high enough by going through kohya (after testing, 16 is not enough, so I put 64)
@n0utis I don't actually play Wuthering waves and I can confirm this is not the same character 😞, also with proper training you can easily remove the character outfit while keeping the Lora weight the same, please check out this Lora for example:
https://civitai.com/models/620442/lynette-genshin-impact-ponyxl?modelVersionId=693574
50 mb size, has proper tags for the character and their outfits.
@Edwardsmith3500 I use kohya, I will do some tests soon to see if I can reduce the weight without losing quality
the images are from danbooru, it's confirmed by using identical tags that the character has a 90%+ resemblance to the original training images.
@n0utis you should not get the dataset from danbooru, but actually from the official character images, collaborations, or in game screenshots
@Edwardsmith3500 I know that this gives a difference in the final rendering. However, I have to provide a lora that can be used on all models of today and tomorrow without giving an unusable result.
With my training parameters, it is possible to merge 2 lora or increase depending on the basemodel used, the weight of the lora to approach the original character.
the worst that could happen for me is that my lora only works on 1 basemodel out of 2 or when a new version comes out for example flux, that it cannot be used on it. All of my models are trained with this in mind.