There's a new version of training called “federated learning”, where you can keep the image on your device the entire time, but you still consent to an update. You tell them all how it should adjust its parameters such that it can better classify your own image. In this case, the image never leaves you or your phone, but the model is still able to use that information to improve itself. However, it is a bit ambiguous how consent would work there.
On April 4th, 2022. See this statement in context.