SARGAN: Spatial Attention-based Residuals for Facial Expression Manipulation

Encoder-decoder based architecture has been widely used in the generator of generative adversarial networks for facial manipulation. However, we observe that current fails to recover input image color, rich details such as skin color or texture and introduces artifacts well. In this paper, present a novel method named SARGAN addresses above-mentioned limitations from three perspectives. First, employed spatial attention-based residual block instead vanilla blocks properly capture expression-related features be changed while keeping other unchanged. Second, exploited symmetric encoder-decoder network attend at multiple scales. Third, proposed train complete with connection which relieves pressure generate face thereby producing desired expression by directly feeding towards end generator. Both qualitative quantitative experimental results show our model performs significantly better than state-of-the-art methods. addition, existing models require much larger datasets training but their performance degrades on out-of-distribution images. contrast, can trained smaller expressions datasets, generalizes well images including human photographs, portraits, avatars statues.