Agreed, results not good for this style of images. I have a model training on a much bigger dataset of image-prompt pairs which should perform better on this.