关于ip-adapter的疑问

[FORRYKIM]

仿照了ip-adapter的解耦交叉注意力机制
ip-adapter中是在每个交叉注意力层后又添加了处理图像特征的交叉注意力层
添加的这一层是需要训练的
请问instantID中没有提到这一层的训练，应该如何理解呢？
The decoupled cross-attention mechanism was modeled after the IP-Adapter.
In the IP-Adapter, after each cross-attention layer, an additional cross-attention layer that processes image features is added.
The added layer is trainable.
How should we understand that InstantID does not mention the training of this layer?