I've sometimes noticed that when switching models mid-conversation, images sometimes don't get sent to the new model and so it loses context. I suspect it might lead to performance degradation with the model essentially "seeing" its own responses hallucinate information from non-existent images. I'm not sure if this is still the case, but I believe I encountered this behavior about two weeks ago. It only happens sometimes and I'm not sure how to reproduce it.
Overall, if this is a bug, it should be able to improve performance and make it a smoother experience for the user.
