Anime Screencap Style LoRA
Making models can be expensive. Do you like what I do? Consider supporting me on Patreon 🅿️ or feel free to buy me a coffee ☕
UPDATE: v3 is here!
v3 should be used at around weight 1 on AnyLoRA ckpt or similar style neutral models.
v3 is trained from ALL my anime screencap datasets, the same I used for all my specific anime models. It should retain "some" styles if you use the trigger words from those, but it's never gonna be accurate to one specific style.
v2 is more "magazine" style and should be used at around 0.3 weight. Trigger keyword is
anime screencap.
How to use LoRA's in auto1111:
Update webui (use
git pulllike here or redownload it)Copy the file to
stable-diffusion-webui/models/loraSelect your LoRA like in this video
Make sure to change the weight (by default it's
:1which is usually too high)
Description
FAQ
Comments (27)
Anything V4.5 has a bunch more Hentai added to vs. V3 which might be why you get so much Hentai unless you included hentai in your training. Did you?
Also, Did you training include series and character names? If so, which ones?
it's a scraped dataset, so there is a lot of hentai and ecchi stuff.
WOOOG :D, well that's what i was trying to do but with a bit of niji journey style to it.
may i ask your help ? it's about your setting for training a style.
sure, contact me.
@lykon add Le Foubre#4650 on discord if you want.
i would like to see a setting and a dataset of yours that i could train. see if it works at all.
as of now, i can't make use of lora and embeddings doesn't train sadly
no need, I'm in basically every discord channels, where there are also people more qualified than me who can help you
So this is like official art tag but with more anime style? Looks pretty good!
How did you do your training? What settings did you use? How did you crop your images, manually or automatically? Can you provide your training images and prompts.
Thanks
settings are irrelevant, Use the default ones with good data. Cropping is useless, use bucketing.
what is bucketing and how do you use it? What about providing your training images and text labels for the images?
@Lykon Could you share your dataset?
@goodfun it takes time and effort to curate a training set get your own don't sponge off others
@psycdoodler that was unnecessary. You don't have to be rude. I'm mostly asking about settings for the training and how to make the training data, what prompts are good to put into the file for each image. Do people bother going that detailed, does it matter that much etc.
@goodfun if that's your concern, this is a style lora, so base that on your base model of choice. Use a model based on booru style captioning (most anime ones) and just use the trigger word like in the examples.
@Lykon What do you mean base model? Loras work on other models too and they are trained on images. This is just too vague to be really useful. What settings, how do you pick images, how many images, how many steps to train, what about overtraining, etc.
@goodfun I'm just saying this lora only has 1 trigger word and doesn't need any particular descriptive word. I managed to use it on many different base models only adjusting the weights. This worked on booru based models but also on dreamshaper which is caption based. I have little doubt it would also work on base sd1.5.
what did you use to train it kohya_ss?
@goodfun yes
What model did you use for the preview images?
anything-v4.5-pruned.ckpt [e4b17ce185]