FoodX-251 is a food image dataset with 251 fine-grained categories, with 120,216 training images, 12,170 validation images, and 228,399 test images. The training images have noisy labels, and the validation and test images have manually verified labels. The categories are fine-grained and visually similar, such as different types of cakes, sandwiches, puddings, soups, and pasta. This dataset can be used to study the field of fine-grained visual classification.