My understanding is that the image data used a decoder-only stage, i.e. mapping images to tokens, basically taking the image textual description instead of the

My understanding is that the image data used a decoder-only stage, i.e. mapping images to tokens, basically taking the image textual description instead of the actual pixels so it can't "see" but can understand the "narration of an image"