search

csxmli2016 / w-plus-adapter

[CVPR 2024] When StyleGAN Meets Stable Diffusion: a W+ Adapter for Personalized Image Generation
Other
119 stars 8 forks source link

RCA's residual setup #4

Open gaoyixuan111 opened 7 months ago

gaoyixuan111 commented 7 months ago

The statement at the end of the WPlusAttnProcessor Class defines the residual connection. Are you defining the initial hidden_states, which is the input from the previous step, as residual, and the hidden_states after W+QKV calculation as the actual functional residual? "The order of key statements is as follows." residual = hidden_states hidden_states = hidden_states + self.scale * wplus_hidden_states hidden_states = hidden_states + residual

Your work is excellent, thank you sincerely for your response.

csxmli2016 commented 7 months ago

The statement at the end of the WPlusAttnProcessor Class defines the residual connection. Are you defining the initial hidden_states, which is the input from the previous step, as residual, and the hidden_states after W+QKV calculation as the actual functional residual? "The order of key statements is as follows." residual = hidden_states hidden_states = hidden_states + self.scale * wplus_hidden_states hidden_states = hidden_states + residual

Your work is excellent, thank you sincerely for your response.

In my understanding about your question, hidden_states = hidden_states + self.scale wplus_hidden_states the latter item ```self.scale wplus_hidden_states``` is the residual in our work.