Twilight Sparkle SDXL LoRA

SDXL version 1.0 Twilight Sparkle LoRA trained on 44 images trained with tags and natural language captions. The model will sometimes fail to generate eye color correctly requiring inpainting and will sometimes diffuse the hair stripe into more of a gradient, adding tags like 'purple eyes' and 'purple hair with a pink stripe' can mitigate this issue but is not always necessary.

Trigger word is 'Twiggy' and sometimes results in a black cat appearing for some reason. happens maybe 1 in 30 times. Prompting with 'Twilight Sparkle' can help but pushes the model towards a more rounded face and 'equestria girl' aesthetic.

The model is highly adaptable to clothing and decent for poses but the lack of a well trained anime finetune is still going to result in some wonky anatomy, all fairly fixable.

Further training is ongoing and I'm up to v6 but V3 was selected for being most adaptable to styles.