The preliminary model gets ~93% accuracy on an "in-the-wild" test dataset of ~10k images pulled from the widely used models, i.e. midjourney v5/v6, SDXL, DallE-3, etc.