ControlNet's zero convolution layer design (initialized with zero weights) ensures that the original model's capabilities are not disrupted during early training, with training costs far lower than training a new model from scratch.