[Question] IPAdapterAttnProcessor2_0 in diffusers is Different from IPAttnProcessor2_0 in AP-adapter source

[MrWH123]

hi ,@xiaohu2015

I get one question on the source in diffusers and in this repo.

Seams in diffusers, each image embedding tokens are processed one-by-one in attention module.

attention_processor.py


this operation means ip-attention is calculated on the first image token, then update the hidden_states. then the second one and so on.

Seams quite strange.

while in the original repo,
attention_processor.py


all image tokens are processed in one attention calculation op, that's the way I think correct.

So could you help to confirm if this is an issue or I made a mistake. Thanks so much