v2 Now model is trained on new dataset and using AdamW optimizer which seemed to produce best quality with this dataset, still has some issues but is as close to OC as I currently can make it.
v1. This model was trained to produce OC of StrunVul, trained on 5 images.