♥️ likes and reviews are always appreciated (especially reviews with images since I suck at prompting myself)
didnt like the other lora on the site (it was kinda fried) so I made my own.
This is a LoCon but it can be used as a regular LORA, all of the sample images were generated with the default lora extension.
You can read more about LoCons here: https://github.com/KohakuBlueleaf/LoCon | https://github.com/KohakuBlueleaf/a1111-sd-webui-locon
weight 1 like god intended
if you want the food wars anime style add "anime_screencap_rindou" to the prompt. It will not do it unless you prompt it.
uniform outfit: plaid skirt, sweater, jacket
you can use the "hair over one eye" and "slit pupils" tags to force her usual look.
Remember, kids: quality over quantity
Description
FAQ
Comments (14)
Wonderful! Will you make Nikumi's Lora?
same answer I gave to the other comment: not enough good art to make it work comfortably. Would either cause it to learn bad poses and would have fucked up hands because most of the images have her holding a knife or something else or it would be hardlocked to the anime style if I only used the screencaps of the episodes.
I've manually drawn over fanart poses in the dataset before for some of my other LORAs and I do not wish to do it again :)
Please make Mito ikumi!
yes
What LoCon Convolution Rank and Alpha do you use?
32 and 32
I have tried lora learning several times myself, but it doesn't work at all...
So I would like to request a character, do you accept such requests...?m(__ __)m
uhh I don't really accept requests but if you tell me the character I'll take a look and see if I'm interested or not.
@gronon Thank you very much!
The character is a hibachi from a work called Musibugyo.
@aaajjjja I don't think there's enough quality art to make a good lora of her
@gronon
I have some images, so is it possible to have them try lora learning with them?
https://14.gigafile.nu/0325-kb6c96ffa81b4358cef154fe834f7d99b
You can download images for reference at the URL above.
please keep them separate from the ones you need.
Thank you in advance...(＞人＜;)
@aaajjjja
\> please keep them separate from the ones you need.
what did you mean by this
trying to train with this dataset would not give good results, there isn't a single image that depicts her outfit correctly. You need images that show the character design elements as you want them reproduced, having skirts of different colors and inconsistent clothing for the same outfit will create severe learning issues. on top of that you also want to minimize NSFW from your dataset, the job of a LORA is to extend the base model and make it able to reproduce a character. Concepts like sex positions, nudity and weird poses are all part of the base model and having those in there without good reason will decrease the overall quality of the model.
to summarize:
you want at least 20 images that show her outfit correctly (honestly can be screencaps from the anime IF you use regularization images)
minimize nsfw from dataset
minimize AI generated images (not necessarily too bad but having a lot of them causes issues)
@gronon Ah, I understand.Thank you very much for your kind words. It was very helpful.
I'll try again !
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.