Queen Chrysalis is a character from My Little Pony: Friendship is Magic. This is her humanized version.
This LoRA took huge amount of time to make (about 8 months of daily attempts), yet I'm not yet content with it. The results might not be consistent, but I want to take a break and do something else for now.
Queen Chrysalis has been trained on vest amount of various tags. Main trigger is queen_chrysalis, with additional chrysalis crown, chrysalis horn, and chrysalis ears. Beside these there are also other tags that were fined tuned to give a specific look for Chrysalis. Continue reading to learn more.
Weight 0.8 - 1.0 is recommended. Main tags sometimes bleed between each other so combining them creates extra weight - lower it to counteract this behavior.
Careful with aggressive negative embeddings, they tend to remove features specific to Chrysalis.
Activation tags
1) queen_chrysalis is the main trigger. Alone it simply creates a woman with long green hair and green eyes.
2) chrysalis crown gives the crown Chrysalis wears in the show.
3) chrysalis horn gives the iconic jagged horn Chrysalis has. Might require increased weight to look good.
4) chrysalis ears adds tiny black ears on side of her head.
Auxiliary tags
1) ears to add human ears (or remove them) since usually she doesn't have them (she has chrysalis ears instead).
2) hair between eyes is a bit overtrained so adding it might require less weight to look good ie. (hair between eyes:0.7)
3) slit pupils or green sclera if you want to adjust the eyes appearance.
4) colored skin and black skin together to give Chrysalis her show skin color.
5) green waist sash is fined tuned to look like that green abdomen thingy Chrysalis has going on in her show design.
6) green sarashi looks kind like green waist sash but around the chest area.
7) eyeshadow gives green eyeshadow.
8) hair over chrysalis horn to have the horn partly covered by hair.
9) torn clothes, torn gloves, torn dress, torn thighhighs are self-explanatory I think.
10) fangs, sharp teeth, grin are also quite overtrained and require less weight to look good.
Misc.
Training human Queen Chrysalis was a massive undertaking because the available images lack consistency. I gathered around 200 images for the dataset and manually edited all of them in multiple ways. For some images I've painted out some parts, for others I've painted them in. I shifted skin colors, moved elements around, and even replaced backgrounds to make the data more varied. Still, it wasn't enough. So for the final push I've used a 3D model combined with AI to create some additional images for the dataset. Yet it's still not as I good as I want it to be. Which is why this is version 0.8 and basically a beta release.
Trained on total of about 140 images divided into 4 separate concepts, with person used as a class. 8 epochs, 64 DIM / 32 Alpha. UNET 2e-4, TEXT 4e-5, Cosine with restarts 3. Base model used for training was animefull_final, but the example images were created on AbyssOrangeMix3.
My third attempt at making a LoRA. Feedback is welcome.