为解决现有图像仿真中动漫风格迁移网络存在图像失真和风格单一等问题,提出了适用于动漫人脸风格迁移和编辑的 TGFE-TrebleStyleGAN(text-guided facial editing with TrebleStyleGAN)网络框架.利用潜在空间的向量引导生成人脸图像,并在TrebleStyleGAN中设计了细节控制模块和特征控制模块来约束生成图像的外观.迁移网络生成的图像不仅用作风格控制信号,还用作约束细粒度分割后的编辑区域.引入文本生成图像技术,捕捉风格迁移图像和语义信息的关联性.通过在开源数据集和自建配对标签的动漫人脸数据集上的实验表明:相较于基线模型DualStyleGAN,该模型的FID降低了 2.819,SSIM与NIMA分别提升了 0.028和0.074.集成风格迁移与编辑的方法能够确保在生成过程中既保留原有动漫人脸细节风格,又具备灵活的编辑能力,减少了图像的失真问题,在生成图像特征的一致性和动漫人脸图像风格相似性中表现更优.
Research on Latent Space-based Anime Face Style Transfer and Editing Techniques
To address issues such as image distortion and style uniformity in existing anime style transfer networks within the field of image simulation,we propose the TGFE-TrebleStyleGAN(text-guided facial editing with TrebleStyleGAN)for anime facial style transfer and editing.This framework leverages vector guidance within the latent space to generate facial imagery and incorporates a detail control module and a feature control module to constrain the aesthetic attributes of the generated images.The images generated by the transfer network serve as style control signals and constraints for fine-grained segmentation.Text-to-image generation technology captures correlations between style-transferred images and semantic information.Experimental results on both open-source datasets and self-constructed datasets with paired attribute tags for anime faces demonstrate that the proposed model reduces the FID score by 2.819 compared to DualStyleGAN,improve the SSIM and NIMA scores by 0.028 and 0.074 respectively.Combining style transfer and editing retains anime facial details while allowing flexible adjustments,minimizing distortion and enhancing feature consistency and style similarity.
万方数据
维普
