In one embodiment, a method includes, by one or more computing systems: receiving one or more non-video inputs, where the one or more non-video inputs include at least one of a text input, an audio input, or an expression input, accessing a K-NN graph including several sets of nodes, where each set of nodes corresponds to a particular semantic context out of several semantic contexts, determining one or more actions to be performed by a digital avatar based on the one or more identified semantic contexts, generating, in real-time in response to receiving the one or more non-video inputs and based on the determined one or more actions, a video output of the digital avatar including one or more human characteristics corresponding to the one or more identified semantic contexts, and sending, to a client device, instructions to present the video output of the digital avatar.