Your image-captioning LSTM is trained with teacher forcing, always fed the true previous word. Teacher-forced validation loss is low, but in free generation one wrong early word often sends the rest of the caption off track.
Adding dropout improved validation loss but lowered BLEU.