Other Kurisu LORAs exist and may be better, this was made for practice. 90 epochs, 1 step for 100 tagged images, 9000 steps total. Variety of poses, costumes, backgrounds and art styles in training data, though majority official art.
Intended features of LORA:
Makise Kurisu in her default outfit as depicted in official art. Long brown/red hair with hair between eyes and long sidelocks, dark blue/violet eyes, general face shape. Brown/tan jacket worn open, off-shoulder and loose with belts and buckles on upper sleeve, striped on end of sleeves and bottom of jacket. White long sleeve shirt with blue highlights on collar and breast pocket on right side, red necktie worn slightly loose. Black short-shorts with white belt and metal buckle cover. Black tights worn underneath shorts. Heeled black ankle boots with flared top.
What helps with consistency towards on-model Makise Kurisu:
Do not tag 'Makise Kurisu', instead the trigger word (kuristina1).
Trigger word (kuristina1) at higher weight, 1.3 seems good.
LORA weight of 1.3 greatly affects consistency, but may alter image quality. Alternatively use suggested prompts below:
Positive prompts of black pantyhose, black short shorts, black ankle boots, long brown/red/auburn hair, dark blue/violet eyes, white pupils, (colours are user choice). Shirt should always be consistent, but if not include white collared shirt, longsleeve shirt, red necktie, breast pocket.
Negative prompts of alternate outfit, alternate hairstyle, hair accessory, short hair, hair tied, denim shorts, sportswear, skirt, sneakers, child, chibi, loli, labcoat (if going for default outfit)
Prompts and a LORA weight of 1.3 very consistent, but should only be used for singular subject image.
EDIT: Been testing recreating her VN sprite look with lora at weight of 1 and no controlnet. Still using trigger word at 1.3 weight. In addition to tags mentioned above, these tags helped with consistency:
[off shoulder::3]:0.3, [brown jacket::3]:0.5, [open jacket::3]:0.5, -- in auto111/forge this disables the tag after 3 steps, letting the lora do the rest of the details after. Means general shape and colour often correct. Pairs well with euler a at ~40 steps for initial gen, hiresfix can use do the rest on a different sampler with more steps.
[blue eyes | grey eyes], (white pupils:1.3), -- this alternates between blue and grey eyes each step, making the colour a bit closer to VN look. Don't try for ringed eyes on generation, save that for inpainting. The lora can try for her VN eyes look, but resolution and focus is too low to look right unless inpainted.
Shirt tucked in, small breasts, long brown hair, white belt:0.3 -- no need to negative opposites.
Negatives of detached sleeves, school uniform, sleeveless, raincoat, labcoat -- jacket still ends up lower on arms than the VN sprite, but matches how it looks in anime/fanart.
Negatives of @_@, star pupils, heart pupils -- simply mentioning pupils in order to make them white causes models to try and turn them stylised. Easier to simply inpaint or fix manually, but wanted to test. Include on face adetailer, unnecessary tokens on generation.
See examples of these settings in 20/10/2024 gallery post. A handful of many, only to show how to make something close to on-model at lora weight 1 through prompting and before any inpaints.
Costume usage:
With no clothing prompted and LORA weight at 1, gives her a jacket and collared shirt, sometimes necktie and shorts, sometimes tights, possibly sneakers instead of boots. Depending on pose will be more consistent to default outfit. Include positive/negative tags above for consistent outfit. LORA weight of 1.3 gives default outfit consistently. Keep LORA weight at 1 for alternate outfits.
Prompt 'labcoat' and negative on 'brown jacket' to get her labcoat look. Works consistently from my testing.
For nude/underwear, have 'jacket' and 'labcoat' in negatives, positive of 'bare legs', 'nude', 'midriff', etc
Official maid outfit wasn't in training data, sorry. See example image of maid cosplay for tags to use in order to approximate.
Has a slight understanding of dr pepper can/bottle. Prompting 'soda can', 'soda bottle', 'holding soda can/bottle', 'dr pepper can', 'dr pepper bottle' should be sort of close, aligning with training data.
Known issues:
Colours not very consistent. Jacket colour can range somewhat wildly, necktie/shirt rarely blue instead of red/white, hair sometimes more blonde than expected. Higher LORA weight and prompts only fixes it seems.
'Hands in pockets' will change jacket shape to one closer to labcoat. This is due to training data association, only images of Kurisu with her hands in her pockets included a labcoat.
Rarely includes signature on bottom left/right, must have missed one when tagging training data, apologies.
Hair can be a little wild even with floating hair/flowing hair in negatives. Up weights on those and take off score_9 if it's a problem, otherwise it's simply down to model trying to fill out empty space.
Eye colour in training data was inconsistent, sometimes official art shows violet eyes other times dark blue. Tag 'dark blue eyes, white pupils' for more consistent to anime look. Include those in adetailer, 'eyelashes' also helps make her look pretty but not necessary. 'ringed eyes' and 'square pupils' should in theory help get official art look but don't, it's too model dependant.