This model was originally something I did for fun after experimenting with a script and failing. GISL was trained on 7k+ images straight from Google; various subjects, lighting, styles, etc, were included. All dataset images were only around 200x100 resolution making the dataset only 58.3 MB including captioning.

This doesn't give the best results when trying specific prompts but it does very well on others. Note, you'll only get good generations if you use 768x768, Hires. fix seems to break it.