JOJO'S BIZARRE ADVENTURE Part 5 Comic Style(hirohiko araki style)
Recommended weight: 0.6~0.9
When used with another LoRA(if needed): use block weight (1,0,0,0,0,0,0,0,1,1,1,1,1,1,1,1,1)
part1,2
https://civarchive.com/models/225796/jojos-bizarre-adventure-part-12-comic-style-or-12
part5
https://civarchive.com/models/207827/jojos-bizarre-adventure-part-5-comic-style-or-5
part7
https://civarchive.com/models/204598/steel-ball-run-comic-style-jojos-bizarre-adventure-part-7-or-7
Description
first release
FAQ
Comments (7)
do you have any examples of adding a stand to this? I saw in the tags used it was an option. love the style!
I added several samples to the gallery. I checked the captions on learning, and prompt "(stand (jojo),2boys:1.3)" sometimes works but low probability..
@markwatkin thank you! either way its great :)
test on my own character lora(frieren, posted below), works very well. could you gives some hints for style lora training?
The training settings are ordinary(mentioned later), but the number of images and the number of steps are large(2856 images and 38120steps), which might be having an good effect.
"conv_alpha": 28,
"conv_dim": 64,
"network_alpha": 8,
"network_dim": 64,
"learning_rate": 0.0005,
"lr_scheduler": "constant",
"lr_warmup": 0,
"max_resolution": "768,768",
"optimizer": "AdamW",
"optimizer_args": "\"weight_decay=0.1\" \"betas=0.9,0.99\" ",
"text_encoder_lr": 0.0001,
"unet_lr": 0.0005,
@markwatkin Around 3000 images seem crazy, but maybe I can figure it out, thanks for the advice. btw, are all these images jojo style? Is it necessary to add some normal style of jojo character to the dataset, which may teach the AI the difference between jojo style and the normal style. If I use the trigger word jojo_style for the style of the target, can the AI understand me if I use the prompt jojo_style next time?
@codexq I cannot say for certain since I haven't tried it myself, but there are LORA models that are being trained using trigger word.
However, I think it might be better to train with overfitting by increasing the number of epochs and steps using only jojo_style images, and then adjust(weight or using LoRA Block Weight), rather than just augmenting the dataset by adding normal images.
Sorry I can't give you good advice.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.