Systems and techniques to dynamically correct a gaze of participants in a video conference are described. A computing device participating in a video conference may determine that a frame of a video stream includes features of a face, extract a portion of the frame that includes a first pair of eyes, and determine that the first pair of eyes are looking in a non-forward direction. The computing device may retrieve, from a database, a stored portion that includes a second pair of eyes that are looking in a forward direction, and modify the frame by substituting the stored portion for the portion in the frame to create a modified frame. The computing device may send the modified frame to a videoconferencing server for distribution to other computing devices participating in the video conference.