Ever wanted to have your own Kwami? Now you can! Presenting: Kwami Style!

Wait, what is a Kwami?

Kwamis are cute little creatures from the show Miraculous: Tales of Ladybug & Cat Noir created by Zagtoon that airs on Disney+ and Netflix, among others. In the show, Kwamis grant powers to their holders, allowing them to transform into superheroes.

How does it work?

For best results, use a prompt with the following format (syntax: <required>, [optional]):

KwamiStyle, <COLOR> <ANIMAL> Kwami [with <DETAILS>]

For instance, you might write: KwamiStyle, blue penguin Kwami to use its most basic form. Alternatively, you can specify details, for example: KwamiStyle, yellow octopus Kwami with blue spots. I recommend keeping the details in that format (first the color then the detail), but you can chain several together. Training data included captions for different colored heads, limbs, tails, patters, etc. so go wild! Though do keep in mind that the more you specify, the less likely it is the resulting Kwami will exhibit all the prompted features.

Tips and tricks

• LoRA weight 1 worked fine for my testing, but play around a little to find what works best for your checkpoint of choice.

• Most SDXL checkpoints should work, but for best results that look closest to the style of the show, I recommend using a checkpoint that can create a 3D-looking style.

• Depending on your chosen checkpoint, you might need to add some negative prompts if you're getting actual animals or other styles than the simple 3D animation of the show.

• Keep colors and animals simple. A yellow snake Kwami will probably work better than a light goldenrod yellow Banana Cinnamon Ball Python Kwami. Though feel free to experiment, of course.

• Due to the nature of the training data (containing all Kwamis shown on screen so far in the show), it works best with bipedal animals, or at the very least, animals with four legs. You can definitely make it work with others, and I've included a few example images of other animals, but it might require trial and error and/or multiple generations to get what you want, and they are more likely to break the general body proportions and design a "real" Kwami should have.

• Because it was trained on screenshots of the show, it likes to generate some random backgrounds with bits and pieces of rooms and other backdrops. If you want to avoid this, add something like white background or in forest to the end of your prompt to give it something to work with.

Parting words

This is my second LoRA. Feedback, comments, and reviews would be highly appreciated. Through testing I already realized a few things I could improve, and if it seems necessary, I might train a new version at a later date. But for now, have fun with it, let me know what does and doesn't work, and don't forget to upload all your cute (or not so cute) creations! I'm really curious to see how it performs on a larger scale and all the little creatures you generate.