Google builds ten-minute AI video from memory, agents and feedback loops

Google Research introduced four connected frameworks for long-form generated video: AI video co-director, CANVAS, A²RD and VQQA. The work treats continuity as a systems problem rather than another upgrade to a single video model.

Four layers manage story, scene, time and repair

AI video co-director selects a creative strategy and delegates work to agents for pre-production, keyframes, video and audio. It operates above Gemini and Veo as an orchestration layer, while a multimodal LLM judges the assembled cut and feeds a signal into the next round.

CANVAS maintains structured memory for characters, locations and objects. A²RD generates segment by segment, alternating extrapolation for narrative progress with interpolation when familiar elements return. VQQA asks visual questions and revises the text prompt rather than editing pixels directly.

Value is moving from the model to production direction

For creators, the important change is where responsibility sits. Costume, spatial and prop continuity can be handled by explicit memory and a control loop instead of a sequence of manually maintained prompts. Google demonstrates a ten-minute film in which A²RD continuously retrieves earlier context.

This looks more like production software than a magical text box. Competitive advantage may shift toward tools that coordinate several models, preserve state and explain why a particular shot was selected.

The benchmarks come from the same research kitchen

Google reports a score of 81.4 on GenAD-Bench, which contains 400 scenarios for 50 fictional brands. LVBench-C contains 120 scenarios and requires an asset to remain absent for at least 10 segments before returning. These benchmarks and demonstrations are tied to the system's authors rather than an independent test of routine production.

The frameworks are research, not an available product with pricing, latency and an editable project file. More agents also mean more generation, evaluation and potential failure points.

Human intervention inside an unfinished scene will determine usefulness

The next test should show whether a director can change one scene without damaging the rest of the film, lock a specific asset and trace the source of an error. Cost and time per usable minute will matter just as much.

If the system survives repeated edits while preserving state across versions, it could remove substantial hidden labor from AI video production. A ten-minute output alone does not reveal how many takes landed on the cutting-room floor.

Lilith's verdict

Google seated planners, memory and a critic around the same film. Their value appears when a director changes scene three and the character in scene eight still carries the same briefcase.