Is there a link to how this works - in terms of nn architecture to combine the embedding of the existing image with the edit instruction?