Google Unveils 10-Minute AI Video That Keeps Characters Consistent Across Scene Changes
Built on Gemini and Veo, the system chains together scene planning, video generation and error checking. The focus is on preserving a character's appearance and a location's layout even as scenes change, and Google released a 10-minute demo video alongside the research.
On September 24, Google published research on generating long-form video that stays coherent from shot to shot even as scenes change. Built on Gemini and Veo, the approach divides planning, production and review among multiple AI agents. Google introduced the components that make it work, Co-Director, CANVAS, A²RD and VQQA, and released a 10-minute demo video as well.
The hardest problem is stitching short clips into a single story. Even when each individual scene looks good, the flow breaks if a character’s clothing changes or the layout of the same room shifts. Google points to a specific weakness in existing pipelines: a small error introduced early on propagates into later clips. The end result is that a human has to go back and fix the same scene over and over.
How characters and locations stay consistent
Co-Director first sets the overall purpose of the video, the narrative structure and the visual mood. Agents with assigned roles then bundle per-scene outlines and visual references into a production plan. From there, the system generates imagery and motion, narration audio and music. The finished video is evaluated by AI, and those results feed back into the choices made at the next production stage.
CANVAS keeps a record of the state of characters, locations and objects, then consults it when the scene changes. The idea is that even when the video returns to a location it hasn’t shown in a while, the earlier appearance carries over intact. Google offered a museum heist scene as a side-by-side example. According to Google’s explanation, CANVAS reduced problems that showed up in the comparison case, such as the thief’s hat disappearing or the layout of the exhibition hall changing.
A²RD generates the video segment by segment while remembering the content and motion of preceding scenes. It adjusts its generation method depending on whether the story is opening up new ground or the appearance of an existing character or location needs to be preserved. In the 10-minute demo, Google showed characters’ faces, clothing details and spatial layouts holding up all the way through.
How errors in the video get fixed
VQQA finds problems by generating questions tailored to the video and having AI answer them. Instead of painting over part of the frame directly, it rewrites the prompt, the instructions given to the AI, and regenerates the video. The resulting candidates are compared against the original request, and the highest-rated video is selected. Google described a case where it corrected an error in which a musician was holding a different instrument from scene to scene.
Google says its own measurements showed improvements in narrative, character and background consistency. In one test, Co-Director’s top quality score was 81.4. The significance of this research lies in the attempt to hand off to AI the work of connecting scenes and making repeated fixes. Whether it actually reduces hands-on effort in real production depends on generation time and cost, together with how much additional human correction is still required.