Some question regarding ref attention processor and IP attention processor.

[xiankgx]

Dear author,

Thanks for your wonderful work.

However, I'm not sure how are you able to combine IPAdapterFaceID and reference attention.
If we look at the code for IP Adapter FaceID, they are replacing the attention processors for both the self attention block and also the cross attention block.

https://github.com/tencent-ailab/IP-Adapter/blob/main/ip_adapter/ip_adapter_faceid.py#L158-L162

where-areas you are applying only the IP adapter Face ID to the cross attention block.

https://github.com/ShineChen1024/MagicClothing/blob/earlyAccess/garment_adapter/garment_ipadapter_faceid.py#L190-L192

Please enlighten me. Thanks.
@ShineChen1024

[xiankgx]

Perhaps you are referring to this instead?

https://github.com/tencent-ailab/IP-Adapter/blob/main/ip_adapter/ip_adapter_faceid_separate.py#L157-L159C36

[ShineChen1024]

MagicClothing/gradio_ipadapter_faceid.py

Line 33 in 8fda45b

pipe.load_lora_weights(ip_lora)