Most of the descriptions generated for the images are likely generic and doesn't contain reference to the actual characters and names. I might explore incorporating some sort of more movie specific description in the metadata of the image to improve queries for specific characters.