People tend to think they are not susceptible to change blindness and overestimate their ability to detect salient changes in scenes. Yet, despite their overconfidence, are individuals aware of and able to assess the relative difficulty of such changes? Here, we investigated whether participants' judgements of their ability to detect changes predicted their own change blindness. In Experiment 1, participants completed a standard change blindness task in which they viewed alternating versions of scenes until they detected what changed between the versions. Then, 6 to 7 months later, the same participants viewed the two versions and rated how likely they would be to spot the change. We found that changes rated as more likely to be spotted were detected faster than changes rated as more unlikely to be spotted. These metacognitive judgements continued to predict change blindness when accounting for low-level image properties (i.e., change size and eccentricity). In Experiment 2, we used likelihood ratings from a new group of participants to predict change blindness durations from Experiment 1. We found that there was no advantage to using participants' own metacognitive judgements compared to those from the new group to predict change blindness duration, suggesting that differences among images (rather among individuals) contribute the most to change blindness. Finally, in Experiment 3, we investigated whether metacognitive judgements are based on the semantic similarity between the versions of the scene. One group of participants described the two versions of the scenes, and an independent group rated the similarity between the descriptions. We found that changes rated as more similar were judged as being more difficult to detect than changes rated as less similar; however, semantic similarity (based on linguistic descriptions) did not predict change blindness. These findings reveal that (1) people can rate the relative difficulty of different changes and predict change blindness for different images and (2) metacognitive judgements of change detection likelihood are not fully explained by low-level and semantic image properties.