I disagree. When training an image-processing AI (traditional, pre-COVID kind) having to produce synthetic data is not the exception. This can be as simple as rotating the image slightly or as complex as making collages and curing the seams (is it a wonder then that GenAI images look like they do?)