NEW VERSION!

V3 NOTES:

Recently came back to this and re-worked the lora. Managed to get some VERY clean frames from the HD/remaster animation. Also manually tagged EVERY image. This took... forever. So, if you want to help me out and buy me some food as thanks, here's this:

buymeacoffee.com/Jairjax

This version should adhere significantly closer to the animation, style, and character. It's around 100 high quality frames! Let me know how your generations turn out. I'll leave the other version up for now too if anyone wants to still use them for whatever reason.

So, onto the fun part. Here are the tags you'll want to use to get her outfit right (accurate to the anime adaptation):

dark elf, navel, medium breasts, blonde hair, long hair, side ponytail, red eyes, cropped jacket

This should accurately cover her upper half. For generations showing the full body or lower half (and hands) make sure to add this:

Blue thighighs, brown boots, gloves, black panties

That's it! just make sure you are using the trigger tags: KURO-SVC-STY-V, anime coloring

Also I imagine 99% of you guys know this by now, but make sure you are using Illustrious checkpoints that are anime friendly. As always, I recommend and use FurryXL Illustrious for my anime based generations.



V1 NOTES:

Scenes of chloe from the Kuroinu H-animated adaptation. Small sample. Specifically made for re-draws and edit on the original animation.

Notes: Use either Euler A or SD3 DPM
It can be used with similar LORA but note, this LORA is only trained on a small sample for re-draws on animation frames from the original adaptation.

It works very well around 0.7 str, in combination with waiNSFWIllustrious. However, if you find the original clothing is too heavily altered, you may need to use 0.8 or 0.9. Too much prompt will cause weird results. The images trained are all in 1080x1920, because that is the format for the animation. This resolution seems to give good results.

V2 NOTES:

V2 simply has a larger pool of images. This means that it's easier to generate a larger variety of images that aren't specifically for re-draws like V1. The intended purpose of re-draws remains the same, for those who do not want to use this LORA like this, I would download V2 instead. The consistency and coherence will be less with more images, but the creativity will obviously be higher. NOTE: ALL PROMPTS WHILE USING THE LORA SHOULD START WITH: Chloe(Kuroinu), Dark Elf

Like V1, use WaiNSFWIllustrious and EulerA. For total clothing changes like in the showcase image above, use LORA str of .5 or .6. Any lower than this and it will differ too much from the original character. You can get away with higher values too, but her original clothing will start to seep in a bit. For specific positions (I.E. squatting or something) that never appeared in the H-animation, the LORA does not have the data, but it is possible to generate. To do this, you'll need to use a vertical resolution. I'll add a reference image as an example. The squat image is a 1080x1920 portrait image. These vertical resolutions are needed when making certain position, like squatting. Use LORA srt of .7 or .8.

FINE TUNING WHILE USING V2: Some things to consider when using V2 is that many of the training images used are close ups of her body or face. These are great references for the LORA, but the generations made can be lackluster, as they will be far too close of shots to look compositionally good in creative generations. Try adding prompts like "full body" to alleviate this issue. This LORA also works very well in image to image. Experiment with creating Dark Elves in various positions with the LORA off, then doing image to image with those image with the lora turned on, essentially converting them to look like Chloe. For this, I would use .8 or so generation str and around .8 LORA str.