The default downloaded model is overtrained, tending to generate the default appearance, and instructions to change this have low effectiveness. To obtain more diverse models, please use other checkpoints provided in the attachment. The lower the checkpoint, the more flexible the model (similarity will also slightly decrease). The training process will be redesigned to strive for better models as soon as possible. The model has been retrained using the same dataset, and the model from the 4000th epoch has been uploaded, but earlier epochs might perform better. On my device, generating a single image takes over 2 minutes, so effective comparisons cannot be made. Based on feedback on this version, earlier versions will be tested and released. Note that I have added multiple checkpoints in the model's zip file. The default version has strong similarity, but if you'd like to try earlier steps for better variation, feel free to choose the steps you prefer. If you achieve better results using other checkpoints, please post here and share your parameters in the comments. Additionally, a model has been trained using 30 publicly available images of Emma Watson.