On the image side, a frozen DINOv2-base was used and embedded only once; on the text side, a uniform 4-layer, width-256 encoder was used, with about 11 million parameters; the text model never had access to images throughout.