They almost certainly generate tokens directly from the image. It would be extremely hard to generate short english descriptions which sufficiently describe the

They almost certainly generate tokens directly from the image. It would be extremely hard to generate short english descriptions which sufficiently describe the images to pass some of those benchmarks.