One label per image. Output is a probability distribution over a fixed vocabulary, produced by softmax over the network logits:

The prediction is $\arg\max_k \hat p_k$; loss is categorical cross-entropy. Top-1 and top-5 accuracy are the standard metrics. ImageNet (1000 classes) was the long-standing benchmark.

The bars on the right show the full distribution, not just the argmax. Image 5 is the same person as image 1; the model splits probability between two classes once the appearance changes.