The uncanny valley effect in face swaps can be particularly disturbing when it comes to eyes. Researchers Wilson, Shic, Jörg, and Jain address this in Towards mitigating uncann(eye)ness in face swaps via gaze-centric loss terms. Their study proposes incorporating a gaze-centric loss during the training process of face swapping algorithms. This methodology significantly improves the accuracy of the gaze direction and reduces viewer discomfort.

• Face swapping advances lead to a distinct perception change, especially around eyes.
• Proposed training process addresses these issues by focusing on eye representation.
• Novel gaze-centric loss term collaboration with a pretrained gaze estimation network.
• Results in more accurate gaze directions and alleviates uncanny perceptions.

This research is vital in enhancing the user experience of digital avatars and could have profound implications for privacy applications and entertainment media. The focus on eyes–a traditionally difficult facial feature to replicate authentically–is a testament to the nuanced understanding required to address the complex challenges presented by realistic human-like images in AI-generated content.