VideoPoet by Google

VideoPoet applies LLM architectures to video generation achieving coherent long-form content with natural language control. Researchers demonstrate language model scaling benefits extend to video domain while creators access precise narrative control. Token-based video representation enables complex editing. Text prompts generate videos with consistent storytelling, character arcs, and environmental continuity. LLM conditioning understands narrative structure producing multi-shot sequences with logical progression. Video editing through natural language instructions maintains temporal coherence. Autoregressive generation scales to longer videos while maintaining quality. Multi-modal capabilities extend to image+text and video extension. Evaluation demonstrates superior narrative understanding vs diffusion models. Free research release includes model architecture and training insights. Substantial compute required for inference. Advances language-video unification research paradigm.