A pretrained captioner (BLIP, base) wrote each caption one word at a time. Select a word to see where its decoder looked while producing it. The photograph is dimmed and greyed so that the gold, which is the attention weight, can be seen; the outlined patches together hold half of the weight on the image. By default the gold is scaled to the selected word's own strongest patch, so every map is visible, but its brightness cannot be compared between words: compare the peak and the number of patches instead. The strongest patch usually receives a few per cent of the attention, and reaches 13% only once across these five captions. A few patches are outlined for nearly every word, including a and with; these attention sinks are real, and they are not looking at an object.