Archon: A Unified Multimodal Model for Holistic Digital Human Generation

Researchers introduced Archon, a pretrained human-centric multimodal model for unified digital human generation.

It combines seven modalities with modality-specific tokenizers and autoregressive training on synchronized data.

Archon also adds memory-efficient video compression and decoding, plus modality-based stepwise reasoning.

The approach aims to improve realism, consistency, and controllability for avatars and talking videos.