Image posts are highly appreciated! You don't even need to leave a review. I just love to see what people are creating with my stuff.

VERSION 3 UPDATE:

-read on for current version or scroll to bottom for changelog

Prompt melusine to get the character/species. Prompt [color] skin or else the generator might default to a human skin tone.

The following tags aren't necessary but shape specific species traits:

-curly ears

-straight ears

-thin tail

-stubby tail

-wings

Melusines, the all female species from Genshin Impact's kingdom of Fontaine. Come in a variety of colors with a few variations of body parts.

Issues:

• Slight/moderate bias towards nudity. Unless you want to risk your Melusines being nude, make sure you specify clothing.

• Has trouble properly using "stubby tail" when prompted. Upon further observation, I noticed during the creation of V3 that V2 also had this same issue. Current V3 dataset has over twice as many images tagged "thin tail" compared to "stubby tail". This stems from the fact that not only do most artists seem to prefer drawing the thin tailed variant, but they generally also make sure to draw the longer, thinner tails to the side of the body or poking around the back, whereas it's difficult/impossible for artists to draw the shorter, stubbier tail from a front angle, and most artists don't draw characters/Melusines from behind.

• Likes to generate the police/guard helmets unprompted, or when trying to prompt a different type of hat. Try to prompt Mamere's beret? You might get it, or you might get a half beret/helmet. Caused by most fanart/existing Melusines in-game being guards. These were all tagged in the dataset, so using "helmet" as negative prompting should mitigate the problem.

Works well with Hassaku-v1.3, AbyssOrangeMix3-A1B, Mistoon_Diamond, and especially with Anything v4.5-pruned if you can still find if/where the real version exists.

Donations accepted through my Linktree if you wanna support me.

CHANGELOG:

Version 1:

Trained entirely on screenshots of in-game models.

Generations are 3d.

Heavy bias towards in-game model clothing.

Uses rounded ears and pointed ears prompts.

V1 description: Melusines, the all female species from Genshin Impact's kingdom of Fontaine. Come in a variety of colors with a few variations of body parts. Due to the lack of Melusine art, this LoRA was trained entirely on screenshots of in-game assets. This causes the default generations to be 3D looking. If you want a more 2D look you'll either need to add specific art style prompts or use a specific art style LoRA. Also due to the lack of Melusine art, this LoRA has a slight bias towards the limited number of outfits that the Melusines in-game wear by default. You'll need to put emphasis on your clothing/outfit prompts to combat this. Due to a lack of... less clothed... in-game assets, it's going to have some difficulty creating anything that requires the chest region of the body to be seen, such as a bikini. Prompting (NSFW) at high power will effectively fill in blanks for... less clothing though.

Version 2:

Trained on half in-game screenshots and half fanart.

Generations can be either 2d or 3d.

No bias for in-game clothing. Heavy bias towards nudity. USE clothing tags. Seriously.

Uses curly ears and straight ears prompts.

V2 description: Melusines, the all female species from Genshin Impact's kingdom of Fontaine. Come in a variety of colors with a few variations of body parts. Changed the rounded ears prompt to curly ears and the pointed ears prompt to straight ears. Felt like better descriptions.

Issues:

• Has a heavy bias towards nudity. Fortunately, version 2 is significantly better at generating clothing. Also, allows for the generation of... NSFW bits. So unless you're looking to see Melusines in a way that definitely isn't shown in the game, make sure you specify clothing. Also, also, allows easier generation of breasts now as well.

• Seems to have an issue with faces when using the "from behind" prompt. I feel like this is because very little of the fanart images artists made showed Melusines from behind.

Extra observations:

• Cranking the power of the LoRA up, (e.g. <lora:MelusineV2:1> --> <lora:MelusineV2:1.3>) seems to reinforce their little mitten/glove hands and results in significantly less human hands/fingers.

Version 3:

Trained on roughly half in-game screenshots and half fanart. 147 images total, 61 artwork, 86 in-game screenshots.

Generations can be either 2d or 3d. Better at 2d than previous versions.

Better at keeping the hands accurate than previous versions.

Slight bias with trying to default hats to the in-game Melusine police helmets. Prompt negative for "helmet" if this issue pops up.

Bias towards nudity. Prompt clothing tags to avoid this.

As with v2.0 onwards, uses curly ears and straight ears prompts.