Review Finds Only One of Google’s Four Video Systems Has Ten-Minute Consistency Results

Four related video systems, one ten-minute consistency test. A review of Google Research’s announcement and its four linked papers finds that only A²RD reports consistency results for generated videos of that length. An AI model judged those results across ten scenarios; no experiment in the reviewed work tests all four systems as one pipeline. That boundary matters if the work is read as evidence that a complete production setup can keep a long story visually consistent.

Google’s September 24 announcement presents Co-Director, CANVAS, A²RD and VQQA as approaches to different parts of long-form video creation. The underlying papers were submitted between March and May. This review compares the announcement and those four papers as captured in the supplied September 25 evidence snapshots; it did not generate videos or rerun their tests. The finding is about what their published evaluations establish, not a claim that the systems could never be combined.

The long-video result belongs to A²RD

Google attributes its ten-minute demonstration film to A²RD. The system makes video segment by segment, using a memory of earlier footage to guide what comes next. It can move a narrative forward or anchor a new segment to a character or setting seen before. Those are different jobs: a story must change over time without accidentally changing the identity of what returns.

Its paper goes beyond showing a film. It evaluates generated videos at approximately one minute on VBench-Long, at three and five minutes on LVBench-C, and in ten ten-minute LVBench-C scenarios. For the ten-minute scenarios, the reported character, environment and object consistency figures come from a multimodal language model judge—an AI evaluator of the video. They are not scores from a reported human evaluation of those same ten scenarios.

The paper also describes human evaluations, but those concern a separate VBench-Long sample, not the ten ten-minute scenarios. Its abstract reports gains of up to 30% in consistency and 20% in narrative coherence across benchmarks spanning one to ten minutes. Neither figure should be presented as a human-judged improvement specifically for the ten-minute films. The duration, benchmark and judge all affect what a result can mean.

An ad is not a ten-minute narrative

Co-Director does test finished video. Its rendered-video evaluation uses four-shot, 12-second ads, and its 81.4 GenAD-Bench result averages advertising criteria. That is evidence about short generated ads, not a measure of whether a character or location stays recognizable through ten minutes of storytelling. The number cannot be placed beside A²RD’s long-video consistency results as though both scored the same task.

The system coordinates creative choices and production steps. Google describes an orchestrator that selects a strategy, narrative structure and visual style, followed by agents that prepare a storyboard and make audiovisual material. A model assesses the assembled result and feeds its judgment into another production loop. That design explains why Co-Director belongs in a discussion of video production; it does not make every evaluation of the system a long-video test.

Co-Director has another consistency evaluation on ViStoryBench-Lite, but it measures static storyboard images. For that test, the authors bypassed the script-generation and video-synthesis modules. It can speak to how visual plans hold together across images. It cannot, on its own, tell a reader how the full production system fares once those images become moving shots.

CANVAS carries its visual plan into shorter video

CANVAS tackles the problem of returning to something seen earlier. It plans shots using records of characters, places and objects, with visual references that can be retrieved when a story revisits a setting. A storyboard is well suited to checking whether a planned character keeps the same appearance or a familiar room keeps its layout. It is not itself a finished video.

That distinction applies to CANVAS’s headline gains. Its authors report improvements over their best-performing baseline of 21.6% for background continuity, 9.6% for character consistency and 7.6% for props consistency. Those results come from storyboard evaluation. Calling them ten-minute video gains would change both the artifact being judged and the claim the experiment supports.

But CANVAS should not be reduced to “storyboards only.” An appendix compares approximately 1.5-minute videos produced by feeding keyframes from each storyboard method into the same Veo image-to-video model. That downstream comparison asks whether better visual plans help the resulting clips. It does not turn CANVAS into an independently evaluated ten-minute video generator. Keeping the appendix in view gives its video evidence its due without stretching it beyond the test.

VQQA improves a prompt, not an entire plot

VQQA works on generated video, but intervenes differently. It asks visual questions about an output, uses an AI model’s critique to revise the text prompt, then generates and assesses candidates. Google describes it as a black-box prompt optimizer: it guides another generation attempt rather than editing pixels in the existing video. Its role is to find a better rendition of a request, not to track every character through a long narrative.

The VQQA authors report absolute improvements of 11.57% on T2V-CompBench and 8.43% on VBench2 over generation without its refinement. Those are benchmark gains for text-to-video and image-to-video work. They support a claim about prompt refinement on the tested tasks; they do not measure whether a ten-minute story preserves its people, objects and settings from beginning to end.

Taken together, the papers describe complementary ways to plan shots, make segments and improve generated outputs. Their tests remain separate. A reader can identify which results concern actual video and which concern a storyboard, then ask how long the video ran and who judged it. That makes A²RD’s ten-minute evidence meaningful on its own terms, while leaving a combined four-system production test as a separate question.