An adaptive model is proposed for gaze estimation based on differential convolutional neural network. The adaptive model incorporates information of head pose and designs a network named Differential Network (DNet) by virtue of differential convolution. The DNet is trained to predict gaze differences in the eyes, calibrate the initial gaze estimations and thus reduce the estimation errors. Through validation on the publicly available dataset Eyediap and comparison with other well-performed gaze estimation models developed in recent years, the experimental results indicate that the proposed adaptive model can estimate gaze directions more accurately under free head movement.