One model to generate digital humans across text, audio, video, and motion

Creating realistic digital humans requires coordinating multiple outputs—speech, facial movement, body motion, video—which typically demands separate specialized models. Archon unifies this with one pretrained model across seven modalities (text, audio, motion, video, and more), trained on 72 diverse tasks. To avoid memory overload from video tokens, they compress video 4× using semantic reparameterization while keeping fine-grained details intact. A stepwise "thinking in modality" approach decomposes tricky cross-modal tasks into clearer intermediate steps. Outperforms or matches prior methods on talking head generation, full-body avatars, and motion synthesis.