Wow, this LoRA turned out really good! Usually for my LoRA training, what I would do is download multiple epochs, test them in a divide-and-conquer algorithm-like sort of way, and try to find the least bad one. This one, nope, download the latest epoch, try generating an image, and it gives me a perfect-looking one on the very first try!

In addition, this LoRA is almost exceptionally good at generating "holding helmet" images, because for some reason there is a lot of fanart of her with her helmet removed and holding it lololol

Would greatly appreciate it if you could read this before posting! https://civitai.com/articles/7060/look-both-ways-before-you-crosspost

Notes:

1. There is only one image of "from behind" (and it's of a toy figure lol), so it might not be accurate. From my personal testing, I find that adding "black bodysuit" prompt helps for accuracy since the back is mostly black

2. Likewise if leggings is not accurate when generating "from below" images, "black leggings" prompt helps

3. Can't generate accurate boots from below (particularly the shoe soles, which are meant to be black but keeps appearing as silver). Seems like some of the images in my training data, including photos of the toy figure and some fanart, have entirely silver boots, whereas the actual boots have black soles. Where can I find close-up pictures of her boots though hmmmmmm...?

4. Weapon doesn't look accurate at all LOL. There's only one image of it in the training data set, and I don't really care about it either way lololol

5. All of the "helmet removed, no headwear" images in the training data set are of ponytails, so if AI automatically generates a ponytail without being prompted, helps to add "ponytail" to negative prompt

Prompts:

go-on_silver, silver bodysuit, silver skirt, silver belt, silver gloves, shiny clothes

boots, silver footwear

dagger, holding weapon

helmet, wearing helmet

helmet, helmet removed, no headwear, unworn helmet, holding helmet

helmet removed, no headwear