Something tells me that image models are sufficiently small, it's easier to just have your differently tuned models sitting side by side so you can easily swap

Something tells me that image models are sufficiently small, it's easier to just have your differently tuned models sitting side by side so you can easily swap and inference on them rather than compiling it into one model.