Your last bit about it being different is a known issue with models that have been trained on purely synthetic data, no?