| Title |
A Study on Multi-angle Face Generation based on U-Net using Diversity Loss |
| Authors |
이영지(Young-Ji Lee) ; 이승호(Seung-Ho Lee) |
| DOI |
https://doi.org/10.5573/ieie.2026.63.3.85 |
| Keywords |
U-Net; Multi-decoder; Multi-view face generation; Diversity loss; σ-head |
| Abstract |
In this study, we propose a multi-decoder-based U-Net network capable of simultaneously generating multi-angle face images from a single frontal face image. Existing GAN and diffusion-based approaches have limited real-time applications due to their complex architecture and high computational costs. The proposed model extracts common features through a shared encoder and trains each decoder to specialize in a specific viewpoint (yaw angle), thereby ensuring the stability of multi-angle face generation. Furthermore, prediction uncertainty is estimated using the σ-head branch in the bottleneck feature and reflected in the diversity loss to prevent collapse and improve the reproducibility of profile faces. Experimental results using the Multi-PIE dataset demonstrate that the proposed model improves average SSIM by approximately 0.008 and reduces average LPIPS by approximately 0.009 compared to the existing U-Net. The improvement is particularly pronounced at profile angles (±60°, ±75°), demonstrating that the proposed architecture is robust across a variety of viewpoints without being biased toward frontal-focused learning. These results demonstrate that the model of this study simultaneously secures efficiency and stability in multi-angle face generation, and are expected to contribute to future research on face synthesis and multi-view image generation. |