You can use it with any model trained in 1.4 or 1.5, even 2.0 512x.

Merge of two LoRAs trained over basic SD1.4, on the same images and parameters with different amount of training steps. Averages to 30 steps/image over 90 training set images, half of them being crops of her face.

This improves over my previous tests by converging on the face more easily on 20 samples. It used to take 100 samples to converge. Her face is the hardest part due to its unusual cartoonish nature, so I tried skewing the training data towards her face, but now it's more heavily biased towards portraits. It's also biased against clothing so use clothing keywords when needed. When doing fullbody, face inpainting is highly recommended. Also, back shots are hard because it always tries to portray her from the front, leading to bad anatomy.