Positioning a vision encoder before the multimodal backbone processes the combined visual and textual data.