Facial Attribute Editing Using Dual-Dimensional Self-Attention Mechanism in Uvit

Currently, encoder-based facial attribute editing relies on convolutional neural networks and Transformer architectures. However, traditional methods often struggle to balance accuracy and efficiency when handling important facial details, leading to less than optimal performance. To tackle these challenges, this paper introduces a Uvit network that uses a Dual-Dimensional Self-Attention mechanism as a facial encoder. Unlike the complex structure of models like encoder4edit, Uvit has a simpler design that improves computational efficiency while maintaining high performance. This simplicity enables faster processing and reduces the computational burden, making it more practical. Experimental results show that the Uvit network achieves high-quality facial reconstruction and effectively manages complex facial details, slightly outperforming existing models across various metrics. Overall, this work represents a significant advancement in facial attribute editing, combining efficiency with high fidelity in detail representation.