Fū (フウ) / Seven Tailed Jinchuriki of Takigakure
First off, i made this is a made this model to test my lora training process, since this character has only a small dataset to work with. The outcome is apparently not so bad so i decided to share the model.
Trained on 83 images with 20 repeats, Recommended weights 0.65-0.9.
The hands are scuffed as always i still don't know what i'm doing wrong with the hands. If you know anything about how to improve hands on models pls let me know.
Clothings can sometime get right or sometime it get a red strap around the crop top. The problem the character is always wearing the giant red backpack almost all the time so there. It seems that AI can't distinguish between the backpack straps with clothings. However by using a camisole tag it provide a better accuracy for crop top.
She has wings but i don't think it'll be a good idea to generate them due to the lack of data.
Basic prompt : fuu, 1girl, solo, hair ornament, orange eyes, hairclip, bare shoulders, collarbone, looking at viewer, aqua hair, fishnets,skirt,groin, detached sleeves,closed mouth, upper body
If you want the backpack try the tag backpack
Description
FAQ
Comments (8)
Best Girl
This loRA is pretty,thx
Can you tell me how to compress loRAs, I can't do it through the Kohyass gui without losing quality
@ForkY I'm not clear what you mean by compress. But if you are saying about having lower model file size then here is the explanation :
The file size is directly related to the network_dim and network_alpha when i first learned about it usually set it at 128 on both which result around 144MB. If you set it both to 64 will get around 72MB.
In one of the guide it said "It's recommended not to set network_dim and network_alpha higher than 64, especially for LoHa. If you want to use a higher value for dim or alpha, consider using a higher learning rate, as models with higher dimensions tend to learn faster."
That mean learning rate must be adjusted if you want to change the network_dim and network_alpha.
May be the problem with your case might be not finding the right balance between those two things.
setting which i copied form the other people which work really well:
unet_lr = 0.0001
text_encoder_lr = 5e-05
network_dim = 64
network_alpha = 64
Ino yamanaka please
@gurwinderb262645 bro there are already 2 LORAs and a 1 TI Embed for her.
You are the best Lora's maker. Can't wait for your next great work ;)
Is there any way to get more Fuu or even Levy From Fairy Tale
is it working with sd 1.5?
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.