About

This LoRA has been trained on Pony Diffusion XL v6. I highly recommend using it along with this model, AutismMix DPO.

The dataset contained Lantern official art, as well as screenshots taken very carefully from her PV and the 7.5 boss fight. This LoRA is able to generate both forms. I'm pretty happy with how this lora turned out. It's been trained hard enough to generate her very accurately, including her clothes, art style, and her overall vibe, but not so hard that you're unable to generate her in other outfits and poses, making it fairly versatile!

I made both Epochs 9 and 10 available. I recommend using epoch 10 with a strength of 1 when generating her in her normal outfit, and a strength of 0.8-0.9 when going for more foreign poses/outfits/expressions. Epoch 9 is only there in case you're still struggling with 10.

Epoch 10 appears to be broken ehe~ I'll retrain this lora eventually.

Prompting

Base Lantern

1girl, blue eyes, short hair, brown hair, hair between eyes, medium breasts

Alt form Lantern

1girl, multicolored hair, pink hair, blue hair, blonde hair, multicolored eyes, blue eyes, pink eyes, hair between eyes, medium breasts

Outfit

gloves, pantyhose, red necktie, black gloves, black pantyhose, black footwear, bandages, jacket, long sleeves, thigh strap, coat on shoulders, black jacket, choker, white gloves, white shirt, boots, black coat, black cape, cape, bandaged arm

Observations / tips

The LoRA is sadly fairly overfit on her in-game appearance due to the fact that the entire dataset is made out of images of her from the game / cinematics. That means that although it's possible to generate her with different outfits, her default outfit's vibe will still remain.

THANKFULLY THOUGH, it's still possible to generate her nude. Somewhat. It really wants to give her a collar / necklace but there's a way to get around it! To do so, I recommend lowering to lora's strength to 0.9 and adding "choker, necklace, collar" in the negative prompt.

The example images are not as flattering as they could be because I couldn't upscale them to fix them up, which is the reason why the full body images look a little odd / washed out.

When it comes to prompting, I highly recommend following this sort of template. This is very specifically for models based on Pony Diffusion XL v6:

Positive prompt:

score_9, score_8_up, score_7_up, rating_suggestive, source_anime, base / alt form tags, pose tags, outfit tags, composition tags

Negative prompt:

score_4, censored, realistic, monochrome, 3D, source_furry, source_cartoon, source_pony, pony, gaping

add "choker, collar, necklace" if trying to generate her fully naked.

Other gen parameters I like:

768x1344, 1024x1024, 1344x768
Euler A
30 steps
CFG 5
FreeU enabled
Lollypop upscaler, 1.5x upscale (2x if generating full body), 0.4 denoising strength, 15 steps

Feedback is welcome and appreciated! It took me a few attempts to get this lora right but I don't know if it's perfect yet and I plan on going over it again at some point to make it better.