This study investigates whether training changes the capacity of visual working memory (VWM). We compared change detection performance for novel and trained polygons. During training, subjects developed familiarity with 8 random polygons. Specifically, 4 polygons from a set of 8 were presented on each trial. After a brief retention interval, one of the polygons changed and subjects judged which one had changed (Exps. 1-2) or whether there was a change (Exp. 3). After 320 training trials, subjects could recognize the trained polygons with high accuracy. In the testing phase, subjects carried out the same task again, only this time each trial might contain all familiar polygons, all novel polygons, or a mixture of familiar and novel polygons. We found that change detection performance improved during training, but the improvement was not limited to trained polygons. We suggest that familiarity of non-nameable shapes plays a limited role in modulating the capacity of VWM.