Reproduces a certain style from the Railgun series. Due to characters being in the training set, the following characters are reproducible (quality of reproducibility is generally worse for characters at the bottom):
Misaka: misaka mikoto
Kuroko: shirai kuroko, twintails
Saten: saten ruiko, hair flower
Uiharu: uiharu kazari, (hair flower:1.1), black hair
Shokuhou: shokuhou misaki, sparkling eyes, blonde hair
Frenda: frenda seivelun, blonde hair, blue eyes
Misaka Imouto: misaka imouto, empty eyes, head-mounted display, goggles on head
Index: index, nun, habit, gold trim, grey hair (other outfits do not seem to work)
Kongou: kongou mitsuko (put hairband into negatives)
Touma: kamijou touma, spiked hair, black hair
Last Order: last order, brown hair, blush stickers, ahoge
The following quality tags are recommended: score_7_up, source_anime. Using higher score tags might introduce a 2.5D style. The following negatives are recommended: empty eyes
The following character's default outfits are more complicated to prompt:
Frenda's outfit: black beret, black shirt, white ribbon, white collar, black pantyhose, white skirt, long sleeves
Shokuho's school outfit: brown sweater vest, pleated skirt, shoulder bag, white thighhighs, short sleeves, elbow gloves, white gloves
See related bounty
Side Notes:
Resizing this Locon is quite feasible. You can reduce it down to ~30MB (with sv_fro and 0.97 as target) to retain the style, but character fidelity will suffer. If you're only interested in the style, not the characters, and want a compact Locon, this is a good option
A portion of the dataset was low-resolution game CG. The style of this CG has finer lines and flatter colors, but it suffers from the low resolution. To invoke this specific style, add the token
game cgand putempty eyesinto negatives
Description
FAQ
Comments (11)
amazing, thank you!
Style + Character pack in less than 200mb bundle? Which also work well in either NSFW & SFW images?
Dude, If I'm not currently in miserable paycheck to paycheck state, I would pay you.
Thanks for your kind words, don't worry about it. Glad you like the LoRA!
Thanks, can this loa reproduce the sportswear?
Hmm, maybe it can be used
school_emblem,shirt,sleeveless,sleeveless_shirt,(((tokiwadai_school_gym_uniform))),(((white_shirt))), tracksuit,(((White sportswear))),shorts,white_shorts,
Very nice sample images.
I am submitting a request to create a Lora in "Bounties." We have sorted the dataset and updated the resolution to 1024x1024 or higher. It may improve the quality of your Lora. Please check it out.
I could give it another try, but I've already incorporated your previous image data into the training dataset, after upscaling smaller images with DAT, and cropping and filtering... (I've posted an older v1.4 to your bounty earlier)
In the next possible version I would probably focus more on improving weaker characters like Index and tagging sub characters, but this would happen at a later date.
Really nice job! I gave the bounty a quick try, but yours seems significantly better. Out of curiosity, what optimizer did you use? I couldn't find any results for "LoraEasyCustomOptimizer.came.CAME(weight_decay=0.05)".
I'm using derrian-distro/LoRA_Easy_Training_Scripts on the dev branch which is wrapping kohya's sd-scripts. It also comes bundled with a custom CAME optimizer which is usually used with the Rex scheduler (also part of Easy Training Scripts). I believe OneTrainer also has a CAME implementation if you're using that.
Could you add Misaka Worst?
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.












