I'm curious why encoder_hidden_state is used in the motion module? The motion module as expressed in the paper is a vanilla temporal attention, not cross-attention.
|
hidden_states = motion_module(hidden_states, temb, encoder_hidden_states=encoder_hidden_states) if motion_module is not None else hidden_states |
|
|
I'm curious why encoder_hidden_state is used in the motion module? The motion module as expressed in the paper is a vanilla temporal attention, not cross-attention.
Open-AnimateAnyone/models/unet_3d_blocks.py
Lines 391 to 392 in f3e014e