This is my first attempt at making a LoRA

I’m pretty new to stable diffusion altogether, and coming here, I was disappointed that there was no model for Millie yet, so I sought to try make one myself, and the results were too good to keep to myself. Of course, feedback is appreciated.

This was trained in 13 epochs with about 150 images in its dataset. Its training model is AnythingElse V4 and it is also the model I prefer for generating images. If you find any other good models for the LoRA, feel free to share.

Activation word and other useful terms

millie is the activation word

yellow sclera, slit pupils, sharp teeth, vibrant red skin, black eyes, perfect eyes are all very useful terms for improving the face. I recommend activating “restore faces” if possible.

Using 1girl is very recommended for improved anatomy.

Good things

I specifically chose images I thought encompassed millie’s design faithfully, and so the LoRA maintains her physique, proportions, and (a bit less reliably) her actual height/size. Of course, you can change all this by prompting it yourself.

Millie (by pure accident) does not usually appear with her tail unless you prompt it with demon tail. This is a happy accident that lets you hide her tail if you think it’s getting in the way or messing up when there’s a lot of props, or if you just don’t like the tail altogether.

Backgrounds can resemble those seen in Helluva Boss if you want some immersion. (The pink and purple striped wallpaper inside I.M.P hq comes to mind)

Millie seems to be pretty good with LoRA’s that entail actions, poses etc.

Bad things

Because of the way I chose the dataset’s images, a lot of generated results appear cartoony akin to the animation style of Helluva Boss. You’ll need to shove the prompt bar with buzzwords like beautiful, masterpiece, extremely detailed etc. if you want something more 3D looking. (Although, millie does seem to be pretty receptive to other LoRAs that change the overall style, so you can use that too if you prefer.)

While the dataset does contain images of Millie’s alternate outfits such as her cat costume, all my attempts at generating them have been unsuccessful. If you find a way, feel free to share.

Backgrounds will be very simple looking unless you prompt otherwise.

The LoRA has a mostly nsfw dataset, so if you are looking strictly for SFW image generation, keep that in mind and prompt/negative prompt accordingly.

I hope you all have fun with this model. I can’t say it’s perfect, but as far as first attempts go, it’s damn good in my eyes.