I chose 53 carefully selected images for my dataset, including from apparent porn along with medical presentations, allowing for versatility without any clear repeats, in as many angles as I could find. Side view is possible since there's a ton of those out there at high quality.

I initially trained for 3,000 steps but the item broke down easily. It still requires multiple gens to get a good hit sometimes but 6,000 steps seems to have given us an acceptable model.