[PonyXL B-LoRA] Kongming コウメイ | Demon Lord, Retry! R | Maou-sama, Retry! R 魔王様、リトライ! R

Kongming from Demon Lord, Retry! R (Maou-sama, Retry! R 魔王様、リトライ! R)

This model was trained on anime screencaps (up to Episode 8) and several AI gens.

Note:
This is a B-LoRA. Use the style LoRA to adjust the intensity of the anime style.

Prompts:-

• Trigger Words:

kongming_olis, 1girl, black hair, very long hair, green eyes

• Outfit:

china dress, red dress, arm strap, red sleeves, panty straps

Optional:
thigh ribbon (does not work very well)

• Negative Prompts (optional):

couch
china (may be required if you put china dress in the positive prompt)
gradient background, simple background

Weights:

Character LoRA: 1
Style LoRA:     0     for a neutral style
                >=~8  for a more accurate anime style and outfit

(Feel free to experiment with the weights.)

Eyes Inpainting: (for anime eyes, use with the style LoRA)

• Manually mask the eyes and set Inpaint area: Only masked

  OR just use ADetailer with one of these models:

  https://civitai.com/models/330727/full-eyes-detection-adetailer OR https://civitai.com/models/150925/eyes-detection-adetailer (set Mask merge mode: Merge)

• Width/Height: 1024

• Padding Pixels: 0-32

• Steps: ~30

• Denoising Strength: no idea (I usually do a X/Y/Z plot with a value 0.2-0.8(+0.1) and then choose the best one.)

• Positive Prompt: add both LoRA and green eyes (and optional looking at viewer)

Known Issues:

• The tag thigh ribbon doesn't work well (and I wasn't even sure how to tag it properly.)

• If you add china dress to the positive prompt, you may need to include china in the negative prompt to avoid bleeding.

Afterword:
The art in this anime is quite bad and inconsistent, so I decided to add some AI gens to the dataset again.
I continued experimenting with eye inpainting and decided to revert to my previous method for preparing portrait images in the dataset. This involves cropping the portraits to include only the character's face, excluding the rest of the head. During inpainting, I used the ADetailer model full_eyes_detect_v1 to inpaint only the character's eyes. This approach seems to work better. As before, reducing the padding pixels might enhance the quality of the eyes.

Hopefully some of you will like this model.