A character's coat can change between two AI-generated shots even when the script says it should stay the same. Google Research described a set of systems on September 24 that track visual details across longer videos.
The work addresses a different problem from making one attractive clip. A story needs returning characters and places to remain recognizable while events move forward. Generating each shot separately makes that harder.
One framework, CANVAS, keeps a structured memory of characters, locations and object states. When a scene returns to an earlier setting, the system retrieves visual references rather than relying on a fresh text description alone.
Another component coordinates a storyboard, keyframes, motion and audio through specialized AI agents. These are programs assigned individual production tasks, with a coordinating system selecting a shared creative direction.
A separate framework, A²RD, generates video in segments while consulting a memory of earlier material. Google shows a 10-minute example. Its length demonstrates a research workflow, not a guarantee that every requested film will remain consistent.
The quality-checking component asks visual questions about an output and uses the answers to revise its prompt. It generates a new candidate rather than directly painting a correction onto existing frames.
Google reports improvements over its comparison systems, but the examples and evaluations come from the researchers. The blog does not establish that the workflow eliminates continuity errors in general use.
The approach treats continuity as information to preserve throughout production.
