增加额外的layer by layer特征，逐层和原始unet求和。那这些额外的特征参数量和unet一一对应？同样是个大模型？

[henbucuoshanghai]

请问这一块细节在哪里？

[guajioao]

4.1中提到"BrushNet utilizes a clone of the pre-trained diffusion model while excluding its cross-attention layers"