Lora of Eida from Boruto
For a more consistent design:
eida, long hair, bangs, blue eyes, very long hair, blue hair, pink hair, multicolored hair, earrings, blunt bangs, two-tone hair, streaked hair,
For your outfit use:
shirt, long sleeves, jewelry, earrings, shorts, star (symbol), nail polish, bracelet, short shorts, black shorts, crescent, star (sky),
With a weight between 0.8 and 0.9.
Clarifications
Recommended weight between 0.65 and 0.8.
For all the photos the model yesmix v15 was used
You can download the photos to see the complete specifications
As most of the images have their clothes, sometimes some items escape in the images even if you don't indicate them, sorry
Description
FAQ
Comments (4)
I'm having an impossible time replicating similar quality results to what you've posted for this. It also doesn't help the images you uploaded for this do not have their PNG exif info embedded so you can't see what possible upscaler or how many steps were used or other data beyond what you listed in the prompts.
I've literally copied everything on the prompts word by word, same seed, same cfg, same sampler, same amount of steps for sampler, same checkpoint, but if I use an upscaler to push a 512x768 to a 1024x1536 (same size as what you have for you posted images), quality goes down the drain and the eyes alone are very poorly rendered.
[last edit] After messing around with denoising values for hiresfix on an upscaler like R-ESRGAN 4x+, I have to place it at a decently high number like 0.45 to 0.6 to get it close to quality without totally manipulating the image to something else.
I uploaded the photos with the PNG exif info embedded but the page doesn't know why it deletes it, there I give you the configuration that the photos had, almost all of them have the same, the only thing that varies is the model between abyss and yesmix, and the round weight between 0.85 and 0.9 for those who have their suit, the others are between 0.75 and 0.8 but there are 2 that I recently realized are from a previous version (the business suit ones), which I have to take but for the rest the information is as follows (this corresponds to the cover image):->
Steps: 30, Sampler: DPM++ 2M Karras, CFG scale: 12, Seed: 499148933, Size: 512x768, Model hash: 0fc198c490, Model: AbyssOrangeMix2_hard, Denoising strength: 0.55, Clip skip: 2, Hires upscale: 2, Hire steps: 20, Hires upscaler: R-ESRGAN 4x+ Anime6B, AddNet Enabled: True, AddNet Module 1: LoRA, AddNet Model 1: eida_Boruto-15(cb895e4f8747), AddNet Weight A 1: 0.85, AddNet Weight B 1: 0.85
Keep in mind that there are 2 textual inversions as well, the easynegative and the badhand
@Sumire I did try a few upscalers like Ultrasharp, Lollypop, Remacri, R-ESRGAN 4x+ and R-ESRGAN 4x+ Anime6B but their initial denoising values were much lower at 0.2 or less and ran with around 10 hireupscale steps. The heavy amount of denoising seems to replicate better results but anything past 0.60 is obviously going to warp the image badly. Also, I had the textual inversions in the negative prompts beforehand when I did some test renders trying to make it work. Odd how the site didn't keep the exif info when you uploaded the images though. One thing I noticed is that this LORA seems extremely stingy in keeping it's core weights into her clothing. If I wanted to give her a different outfit entirely I'd have to push something like "(((white bikini:1.7)))" when I was using YesMix with the lora around 0.8 like you suggested in the description. I forgot to try using AOM checkpoint for her as I only used Yesmix v15 on my initial batch.
It's not a bad LORA, could use more training and updates down the road when more pictures and media for Eida comes out when the anime and manga return in September this year. I should've kept the generations both bad and good so I can upload them here, I had but I deleted them as I sort of already moved on with this LORA.
this is just twilight sparkle
???
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.