A.K.I. from Street Fighter.
Illustrious / NoobAI :
v1 :
Trained on illustriousXL_v01
Weight 1 - 0.8
Trigger : AKI, short hair, White hair, hair over one eye, red eyes, mole under mouth, makeup, tassel earrings, chinese clothes, black dress, snake print, see-through sleeves, Long sleeves, wristband, chain around arm, see-through bridal gauntlets, side slit, hip vent, pants, pelvic curtain, black footwear, claws, claw (weapon)
Would be awesome if u post a pic ;)
Description
FAQ
Comments (7)
A really well-done LoRA for Aki. It offers a generally faithful representation of the character. However, her costume is quite complex, and the LoRA rendering varies from image to image. You will therefore often need to adjust the tag weights to obtain a representation close enough to Aki's exact costume. I am going to post a set of images to illustrate my points.
hey glad u like it! but yes the costume varies. i guess i mixed in to much fanart (which isnt perfect but much better them my images). especially the snake pattern gets lost and the low pants.
@BloodySunkist For the "snake print" pattern, i had to associate it with the relevant garments. Therefore, I used the terms "snake print dress" and "snake print pants" rather than simply "snake print" and then other Danbooru garments tags. The "see-through" tag also caused problems with her "long sleeves" one. Most of the time, I ended up with patterns on her long sleeves because of the "snake print" tag that bled onto her long sleeves if not using this LoRA at a weight score of "1". At the ned of the day i ve chosen to get a "grid print" pattern on her long sleeves except for the final image. That said, the main consistency issue with this LoRA when it comes to Aki's clothings lies in the representation of her gauntlets/bracers/wrist bands. The results will never be consistent from one image to the next. To mitigate this problem, I preferred to use Danbooru tags rather than those used for LoRA training, as they were too generic to accurately represent Aki's gauntlets/bracers/wrist bands. Increasing this LoRA weight helps but at the cost of pose flexibility.
@antarek You´re right about the tags. But i didnt want to use tags that were already associated with the char, built into the checkpoint. And sometimes its a hassle to find the right tag as a non native english speaker. that should not be an excuse ;). And now that u mention the bracer/wrists, the dataset images were indeed mostly different in that area. But i like to use fanart so i get "a unique version of the char" same with my tifa and lol chars. i find it boring to just use official art. it feels even less artisic then ^^. it just feels like stealing.
@BloodySunkist That's understandable. I tend to stay as true to the original character as possible, unless otherwise specified. Especially if I'm using them in one of their official outfits. And that's what most people who ask me to create a picture set are looking for: a character as faithful to the original as possible. And your LoRA is a perfect tool to achieve that objective. Thanks again for your work on it. 👍
Oh and also her "mole under mouth" tag is a real issue for as far as 99 times out of 100 she ll get more than a single mole under her mouth. Especially if using ADetailer no matter if you place the "mole under mouth" tag in the base prompt or the ADetailer one. Meaning that you ll often be forced to use inpaint to get rif of those extra moles under her mouth. And this isn't something that i want for as far as i try to post pics that people ll be able to reproduce simply by using my prompt and parameters. Meaning that i don't want to have to use img2img or inptaint for exemple. 😉
@antarek Yeah stuff like that is the checkpoints/illustrious fault. u have to train for ages so that only one appears. and random is builtin into ai, so so get more then one this or that way. best is dont use the tag and inpaint it.
Details
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.