Google Research studies multi-agent direction for longer AI videos
The work combines planning, state tracking and visual refinement to reduce character drift across generated shots.
THE STORY IN THREE POINTS
- Google described the research on September 24.
- The framework orchestrates Gemini and Veo for multi-shot narratives.
- The work targets continuity errors and failures that propagate between stages.
The research
Google Research described a multi-agent approach to longer video narratives, built as an orchestration layer above Gemini and Veo. The work treats continuity as a planning and state-tracking problem across multiple shots.
The post presents Co-Director and related frameworks for coordinating creative decisions and visual refinement. It reports improved narrative consistency in evaluations. The announcement concerns research methods; it does not establish a generally available film-production product.
Editorial context: consistency is a storytelling requirement
A sequence can contain individually convincing clips while losing the identity of a character or the layout of a room. A viewer then has to reconstruct what happened instead of following the intended story.
An evaluation could specify a short narrative and identify the details that must persist across cuts. Reviewers could separately assess visual continuity, whether events advance the story and how much manual repair the sequence requires.
Editorial outlook: what to watch
The useful next comparisons would report editing effort and failures across full sequences, including longer and more complicated briefs.
Creative control also matters: a director should be able to revise one decision without unexpectedly changing everything else. That would test whether orchestration supports a practical production workflow.
Sources: Google Research: Automating coherent long-form video generation ↗
