This was made after my first attempt at training a LoRA on my friend, I think it went well. I’m currently working on improving the training data to get better results.
For my next attempt I want to improve the eyes because they were pretty inconsistent. Can anyone show me to some good resources for that?

Technical Information

Here’s the approximate prompt, as it varied depending on what facial expression I wanted:

masterpiece, best quality, ({disney|pixar}:1.1),  snapgirl, 19 year old girl, young girl, (detailed eyes:1.3), dark blue hair, short hair, (bangs:1.2), round face, black choker, candid, small boobs, (cleavage:0.7), casual clothes, (on a plane:1.2), eyes focus, perfect eyes, looking to the side, low angle shot
Negative prompt: easynegative, negative_hand-neg, badhandv4, Unspeakable-Horrors-24v, male, man, boy, noisy, jpeg artifact, piercing, piercings, metal, (selfie:1.5), gloves, bad teeth, goofy teeth, real, photoreal, photo, hand, hands

Note that snapgirl is my custom LoRA, you won’t be able to find it anywhere until I release it (if I ever do). The model used was disneyPixarCartoonTypeA_10, which you can find on CivitAI. I also used the Shirtlift LoRA for the last image.

  • SquishyPillow@burggit.moe
    1·
    1 year ago

    Are you inpainting the eyes at all? Eyes often get screwed up due to autoencoder artifacts. Some models that use a custom autoencoder like AnythingV3,4,5,etc. don’t have this problem, probably because their training data contained lots of images of people with large eyes. You could try training a custom autoencoder, or use a model with one.

    Or just inpaint the eyes, and make sure that the inpaint area is set to “only masked”.