Dunhuang murals, as invaluable cultural heritage, hold distinct artistic significance. Transforming them into line drawings aids in preservation and creative reuse. However, current deep learning-based style transfer methods often suffer from blurred lines, style inconsistency, and poor structural fidelity. To overcome these challenges, this paper proposes DMLT-GAN, a GAN-based model integrating multi-level convolution, asymmetric cycle consistency loss, and gradient loss to enhance the stylistic fidelity and line clarity. Experiments demonstrate that DMLT-GAN achieves superior visual quality ...
