The interesting thing is that it was trained only on 6 tasks and it was still able to generalized to unseen tasks, this is only possible thanks to the frozen SD

The interesting thing is that it was trained only on 6 tasks and it was still able to generalized to unseen tasks, this is only possible thanks to the frozen SD parameters and the initialization of the ControlNet model with the weight of the original model (+ zero convolution). Probably training the model on more than 6 tasks will even further increase its in-context learning capability.