基于几何-语义约束与扩散增强的线描生成OA

As a historical relic of China,colored pottery has important artistic and cultural research value.Aiming at the problems of structure loss,line distortion and detail blurring,which are common in the generation of colored pottery line drawings,a two-stage high-fidelity reconstruction model,GS-CycleDiff,is proposed.Based on CycleGAN,the geometric loss and semantic loss are designed.The pseudo-depth map generated by MiDaS monocular depth estimation is aligned with the original photo depth map respectively to ensure the coherence of the line drawing at key geometric structures;and with the help of the CLIP model to extract the semantic features of the image,the correspondence between the constrained generation result and the original image at the level of cultural symbols is carried out by minimizing the distance between the CLIP of the input photo and the generated line drawing.Subsequently,the preliminary generated line drawings are fed into a lightweight diffusion denoising network,which suppresses background noise and strengthens line clarity through multi-step iterative denoising and detail enhancement.The experimental results show that the images generated by GS-CycleDiff significantly outperform the traditional CycleGAN model and other comparative models in terms of line clarity,geometric structure,semantic consistency,and overall visual realism,and are capable of generating fine line-drawing images in multiple styles and complex backgrounds.