Void-aware GAN for modeling structural aspects of creative intent in Chinese ink painting

Chinese ink painting is characterized by a delicate balance between expressive brushstrokes and the intentional use of empty space, where void regions play a fundamental compositional and aesthetic role. This paper introduces VoidGAN, a novel generative adversarial network designed to synthesize Chinese ink paintings while explicitly preserving the structural distribution of void regions. To achieve this, we propose a void-aware learning framework that integrates a U-Net–based generator enhanced with spatial and channel attention mechanisms, along with a dedicated void-aware loss to enforce the preservation of blank regions. A statistically grounded definition of void regions is established using grayscale intensity analysis, supported by Gaussian Mixture Modeling and Hartigan’s Dip Test, ensuring reliable extraction of void masks. Additionally, perceptual loss based on a pretrained VGG19 network is employed to maintain structural and stylistic consistency. The model is trained on 1,754 images and evaluated on 438 samples from the Chinese Landscape Painting Dataset over 1,000 epochs. Experimental results demonstrate that VoidGAN outperforms baseline methods, achieving a Fréchet Inception Distance (FID) of 28.5, a Void Intersection over Union (Void IoU) of 0.82, and lower reconstruction error as measured by RRMSE. Qualitative analysis further confirms that the generated images preserve both the spatial distribution of void regions and the stylistic characteristics of traditional ink paintings. These results highlight the importance of explicitly modeling void space in generative art systems and demonstrate the effectiveness of the proposed approach for culturally grounded image synthesis. Future work will explore adaptive void modeling and extensions to more complex artistic domains.