Google’s AI video co-director tackles the long-standing problem of keeping stories coherent

Google researchers have introduced a multi-agent system designed to generate minutes-long videos without letting characters, props and locations quietly mutate between shots. The work matters because it treats AI video as a continuity and planning problem, not merely a race to make one impressive clip.

Google DeepMind Watch analysis

What happened

In a research post published on 24 September, Yale Song and Yiwen Song describe an AI video co-director built as an orchestration layer over Gemini and Veo. The system plans stories, generates visual and audio assets, evaluates assembled footage and feeds the results back into another optimisation loop.

The research combines four frameworks. The co-director uses a multi-armed bandit to explore creative strategies, narrative modes and visual styles. CANVAS maintains structured records of characters, locations and object states, so a returning character should keep the same clothes and a revisited room should retain its layout. A²RD generates video segment by segment while switching between extrapolating the story and anchoring new footage to remembered visual context. VQQA creates visual questions, uses model critiques to refine prompts and selects the strongest result across the generation process.

Google’s examples compare CANVAS with direct Gemini-3.1-Pro generation and another multi-agent system on a museum-heist sequence. The researchers say the new framework better preserves character identity, spatial geometry and recurring objects. They also present a ten-minute generated film as a demonstration of A²RD’s attempt to maintain continuity across longer temporal gaps.

Why it matters

The useful shift here is architectural. Current video generators can produce striking individual shots, but long-form work exposes the joins: a hat disappears, a prop changes shape or a room develops a new floor plan between cuts. Those errors are not cosmetic when the aim is a story rather than a sequence of loosely related postcards.

Google’s approach gives the system a persistent world state and a way to judge its own output before selecting a final cut. That could reduce the manual prompting and editing currently needed to keep generated narratives on the rails. It also suggests that future video tools may depend as much on orchestration, memory and evaluation as on the underlying video model.

Our read

This is one of the more credible directions for AI video because it targets the unglamorous failure that viewers notice immediately: the story forgets what it just showed us. The reported results are promising, but the supplied research post is not a substitute for independent replication, production-scale testing or a clear account of generation cost.

Still, the central idea is sound. If AI video is going to move beyond short demonstrations, it needs something resembling continuity editing, a production bible and a slightly obsessive script supervisor. Google is now trying to build those functions into the generation loop.

What to watch

• Whether the frameworks appear in a public research paper, code release or usable product.
• Whether independent tests reproduce the reported gains against strong video-generation baselines.
• How much compute and generation time the repeated planning, judging and refinement loops require.
• Whether persistent visual memory improves narrative coherence without making stories rigid or repetitive. Sources and evidence: Google Research Blog post published 24 September 2026, “Automating coherent long-form video generation”. The report describes the researchers’ own frameworks, demonstrations and evaluation claims.

Discussion spark: Should long-form AI video be judged mainly by visual realism, or is reliable continuity the more important breakthrough for making it genuinely usable?

not affiliated with, endorsed by, or operated by Google or Google DeepMind