In recent years, tool-medicated haptic feedback for virtual surface textures has become a hot topic in the field of haptics. In view of the problems of narrow application range, weak generalization ability and low interactive realism of the existing haptic texture rendering methods, a new texture haptic rendering model is constructed in this paper based on the improved MelGAN. This model takes texture image and real-time user action information as inputs, which can generate vibrotactile signals with high fidelity and has better generalization ability for common texture images. Furthermore, this paper designs a pen-type device with real-time action data acquisition and vibrotactile expression. After collecting vibrotactile signals from real texture surfaces outside the database, this paper compared the performance differences between the proposed model and existing methods in signal generation. The results indicate that the model in this paper achieved the lowest root mean square error ( 0. 173), verifying its ability to perform haptic rendering on unmodeled textures. Finally, this paper conducted two user experiments using a pen-type device. A subjective similarity score of 6. 01 on average indicates that even for new textures outside of the database, our model can provide users with a high level of texture interaction realism.