preprint · arXiv
An Interactive Conversational 3D Virtual Human
Richard Shaw, Youngkyoon Jang, Athanasios Papaioannou, Arthur Moreau, Helisa Dhamo, Zhensong Zhang, Eduardo Pérez-Pellitero
In plain English
ICo3D is an end-to-end interactive digital human: a photorealistic full-body Gaussian-splat avatar captured from multi-view video, with an audio-driven head model merged seamlessly onto a dynamic body, plus speech recognition, a small quantised language model and text-to-speech so it holds a spoken conversation. The authors report the viewer running at roughly 70 FPS on an RTX 2080 and 105 FPS on a laptop RTX 4090, with the head model leading prior talking-head methods on lip-sync confidence while trailing on some perceptual metrics. It is a useful reference architecture for teams who want a conversational avatar that renders locally on consumer hardware instead of streaming frames from a data centre; the authors note the version of record is published in the International Journal of Computer Vision.