[PonyXL B-LoRA] Saberia | The Healer Who Was Banished From His Party, Is, in Fact, the Strongest | Party kara Tsuihou sareta Sono Chiyushi, Jitsu wa Saikyou ni Tsuki パーティーから追放されたその治癒師、実は最強につき

Saberia
from
The Healer Who Was Banished From His Party, Is, in Fact, the Strongest
(Party kara Tsuihou sareta Sono Chiyushi, Jitsu wa Saikyou ni Tsuki
パーティーから追放されたその治癒師、実は最強につき)

This model was trained on anime screencaps (up to Episode 4) only.

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

Prompts:-

• Trigger Words:

saberia_olis, 1girl, red hair, short hair, purple eyes

• Main Outfit:

fur scarf, cleavage, crop top, brown shorts, white belt, bridal gauntlets, garter straps, thigh boots

• Negative Prompts (optional):

dagger, knife sheath
blurry background, grey sky

Weights:

Character LoRA: 1
Style LoRA:     >=0       (Her character and outfit are simple, so 0 is fine, IMO.)
                ~0.8-1    for a more accurate anime style

(Feel free to experiment with the weights.)

Afterword:
I've always struggled with eye details in my character LoRAs, so I experimented with eyes inpainting again. This might be common sense, but I realized that the closer the inpaint mask matches the dataset images, the better the results. At first, I cropped only the eyes from the portrait images for training and masked just the eyes during inpainting. However, the results weren't good enough, likely because I used with a square resolution for inpainting, while the eyes were trained in a landscape resolution.
After some unsatisfactory attempts, I decided to include the eyebrows and mouth along with the eyes in each image, allowing me to make all images square, which I then processed to a 1024x1024 resolution. During inpainting, I masked the eyes, eyebrows, nose and mouth together and then inpainted with a 1024x1024 resolution. The results were a lot better than what I did before. Reducing the padding pixels further improved the results, though this required a lower denoising strength.
I'm not sure if this is the right way for improving eye details in a character LoRA, but at least it works for me. The only drawback is that over half of the training steps were spent on training the eyes, which is probably why the anime style ended up a bit overcooked.

Hopefully some of you will like this model.