Misimplementation of Spatial-attention? about moore-animateanyone

[somaonishi]

Sorry, my mistake. Your code is also correct.

[somaonishi]

I've reopened this because there's one aspect that concerns me.

The distribution after projecting $Q$, since it does not include the embedding of the ref_image, might deviate slightly from the intent described in the paper. I'm unsure about the extent of this impact. Has this been considered? (It's possible that it might not have an impact since it may not be involved in gradient calculation.)