Given a reference image and the corresponding prompt, the keyboard or mouse signal, we transform these options to the continuous camera space. Then we design a light-weight action encoder to encode ...