Normally when you train a LoRA with a trigger word, you want to choose a word that the base model doesn't recognize or handle well. Otherwise, you can create a conflict between the way the base model interprets the word and the way the LoRA interprets the word. But for this LoRA I set out to deliberately cause a clash between the LoRA and the base model. And what better trigger word to try this on than the most popular prompt term across Civitai: 1girl.

I trained the 1girlDistortionU LoRA on 176 high quality images of various animals from Wikimedia Commons with manual captions using 1girl as the trigger word. At low weights, it tends to add an animal companion to the image or add features like animal ears, wings, or a tail. At higher weights, it will replace the subject with an animal.

I trained the 1girlDistortionL LoRA on 112 high quality images of various plants and fungi from Wikimedia Commons with manual captions using 1girl as the trigger word. At low weights, it tends to add flowers to the image or move the setting outdoors. At higher weights, it will replace the subject with a flower or plant.

I trained the 1girlDistortionLU LoRA on 288 high quality images of various animals, plants, and fungi from Wikimedia Commons with manual captions using 1girl as the trigger word. The training images were biased toward animals, so this LoRA generally produces animal-like images but with more floral imagery than 1girlDistortionU.

To generate the example images, I went to some of my favorite SD 1.5 models and found a 1girl prompt with a lot of reactions. I reused the prompt with the same model, but with a 1girlDistortion LoRA applied.

Finding the right weight for a given prompt can be a little tricky, but the results can be truly disturbing. So if the waifus you've been generating lately are starting to feel a little vanilla, give this LoRA a try.