Uploads an image and returns a single-paragraph, copy-ready English description. Uses an ensemble of captioners + deterministic post-processing to improve accuracy.