Semi‐Interactive Thangka Floral and Leaf Element Image Generation Network Based on Hand‐Drawn Sketches

Regong Art is one of the world's intangible cultural heritage, and Thangka, as an important form of expression within it, is created in strict adherence to the unique compositional principle of the ‘Three sutras and one commentary’. Current generation methods, when applied to the Thangka floral and leaf dataset, struggle to meet its specific compositional and color restoration requirements. To address these issues, the adversarial ResNet with omni‐dimensional dynamic convolution‐pixel2stylepixel (AROD‐PSP) model is proposed. This model uses hand‐drawn sketches to guide the generation of floral and leaf elements in Thangka, restoring its artistic features. A lightweight improvement to the encoder network is made; omni‐dimensional dynamic convolution (ODConv) is introduced to enhance feature extraction capabilities and increase result diversity; a latent code discriminator is added to ensure that the generated images align more closely with the compositional rules of the dataset; and a colour correction module is employed to adjust the colours, making them more similar to the original images. This approach successfully achieves intelligent design and visual creation tasks for Thangka floral and leaf elements. In comparative experiments, the model achieved second‐best results in quantitative analysis and the best performance in qualitative analysis. The model's generalization ability is demonstrated using two public datasets.