Recent works in the fields of computer vision and natural language processing have enabled the recognition and identification of objects in images, generating automatic descriptions. Despite these advancements, the main research in this field is primarily related to the English language, requiring some adaptation when dealing with other languages, such as Portuguese. One of these methods is the translate-train approach, which involves translating the training dataset into the desired language. However, there are various translators with different levels of effectiveness available. The primary objective of this work is to evaluate the behavior of image captioning models when trained on datasets translated into Portuguese by different automatic translators, both quantitatively (cost, training time, metrics on the test set) and qualitatively (comparative evaluation form, error analysis). The results indicate that it is possible to obtain valid automatic descriptions in Portuguese from image captioning models trained on translated datasets, and that more robust translators produce more meaningful descriptions.