A CNN-based interactive contrast enhancement algorithm, called IceNet, is proposed in this work, which enables a user to adjust image contrast easily according to his or her preference. Specifically, a user provides a parameter for controlling the global brightness and two types of scribbles to darken or brighten local regions in an image. Then, given these annotations, IceNet estimates a gamma map for the pixel-wise gamma correction. Finally, through color restoration, an enhanced image is obtained. The user may provide annotations iteratively to obtain a satisfactory image. IceNet is also capable of producing a personalized enhanced image automatically, which can serve as a basis for further adjustment if so desired. Moreover, to train IceNet effectively and reliably, we propose three differentiable losses. Extensive experiments show that IceNet can provide users with satisfactorily enhanced images.