Trained with 67 images on the AAM - AnyloraAnimeMix checkpoint. The samples were generated with AAM.
Recommended settings:
Sampler:DPM++ 2M Karras / Euler a - 35 steps
CFG: 7
Resolution: 512x768, 768x512, (works with most common SD1.5 resolutions)
Lora weight: 0.8
HighRes: 4x_NMKD-Siax_200k, upscale by 2x, 25 steps, 0.35-0.6 denoising
Basic prompt:
<lora:redacheron_SD:0.75> redacheron, 1girl, long hair, white hair, red eyes, medium breasts, sleeveless shirt, black shorts, cleavage
- Additional tags of interest: "red flower", "tattoo" or "bodypaint" (for the red lines on her body), "holding sword"+"katana" (might take a couple of gens before it get's her katana right)
Additional notes:
Make sure to describe the background/setting in detail, otherwise you will get simple backgrounds.
I used the FastNegativeV2 embedding by Lykon for the samples.
Works well with other loras but you might want to drop the weight of the Acheron lora to 0.6.
Most of the training images didn't have the red flower over her eye, so prompting for it is kind of hard. You can add the flower via inpainting.
Prompt for inpainting the red flower:
<lora:redacheron_SD:0.8> redacheron, one red flower, white hair
Inpaint area: only masked
Sampler: euler a - 20 steps
Resize to: 768x768
Denoising strength: 0.8 (if you're using an inpainting model you can crank it up to 1.)
Tested models:
AAM - AnyloraAnimeMix v1
EtherBluMix v6
Yuzu v1.1
BloodOrangeMix