AI video is getting much better at producing impressive individual clips.
The harder problem starts when those clips need to become a coherent story.
Characters change appearance. Products subtly shift shape. Locations lose details. Objects disappear and reappear differently. The longer the sequence runs, the more opportunities there are for continuity to break.
Google Research is now tackling that problem through a new approach to coherent long-form video generation. Published on September 24, 2026, the research introduces four frameworks: Co-Director, CANVAS, A²RD and VQQA.
Together, the research explores different parts of the same challenge: how to maintain visual and narrative consistency as AI-generated videos become longer.
As part of its A²RD research, Google also demonstrated a ten-minute AI-generated film designed to test whether characters, environments and visual details could remain stable over much longer sequences.
The ten-minute runtime is the attention-grabbing part. The more important development is the architecture behind it.
Instead of relying on a single prompt and generator, Google is building an orchestration layer around models such as Gemini and Veo, with different systems responsible for planning, memory, generation, evaluation and correction. Google says this approach also inherits safeguards such as SynthID watermarking from the underlying models.
Why long-form AI video still breaks
Most current AI video tools are strongest when the task is short.
An eight-second product shot or cinematic office scene can look convincing. The problems become more obvious when the story continues across multiple shots.
Google describes several versions of this problem, including semantic drift, feature drift and cascading failures. A character’s clothing may change between shots, a location can slowly lose its original structure, or an error created early in the process can affect everything generated after it.
For marketers, these issues are more than visual imperfections. A short social clip might survive one strange background detail. A product explainer cannot keep changing the product, presenter or environment without looking unreliable.
That is why the next challenge in long-form AI video generation is not simply better resolution or motion. It is continuity.
How Google approaches coherent long-form video generation
Google breaks the problem into four foundational areas. The systems are part of the same research direction, but each targets a different bottleneck and is evaluated using different benchmarks.
| Framework | What it does | Why it matters |
| Co-Director | Plans creative direction and coordinates multiple production agents | Keeps the broader story and creative strategy aligned |
| CANVAS | Maintains structured visual memory for characters, objects and locations | Reduces visual and environmental drift |
| A²RD | Generates longer video segment by segment using multimodal memory | Maintains context across extended sequences |
| VQQA | Generates visual questions, critiques output and iteratively refines prompts | Helps correct compositional and consistency problems |
The broader shift is important. AI video competition has largely focused on which model can create the best clip. Google’s research asks a different question: what if the generator is only one part of the production system?
That is similar to the move toward multi-agent systems we covered in Sakana Fugu’s multi-agent orchestration, where complex work is split across specialized components rather than handed to one model.
Co-Director turns video generation into a managed production process
Co-Director handles the higher-level creative structure.
Google describes an Orchestrator Agent that selects combinations across creative strategy, narrative mode and aesthetic direction. A Pre-Production Agent turns that direction into a storyboard, while specialized agents create keyframes, video and audio.
An evaluator then critiques the compiled result and feeds a reward signal back into the system so later attempts can improve.
This matters because many AI video workflows still depend on a human repeatedly generating a clip, spotting the problem, rewriting the prompt and trying again. Co-Director explores how more of that orchestration can be automated.
The idea mirrors the logic behind AI marketing automation: the valuable system is not simply the one that produces an output, but the one that can evaluate what happened and improve the next attempt.
CANVAS gives AI video persistent visual memory
CANVAS addresses one of the clearest problems in longer AI video: remembering what the world is supposed to look like.
The system maintains structured representations of characters, locations and object states as the story develops. When something appears again later, CANVAS can retrieve earlier visual anchors instead of recreating the scene from scratch.
That has obvious relevance for branded content. If a spokesperson appears in scene one and scene nine, the person should still look like the same person. The same applies to products, interfaces, props and recurring locations.
We already see this limitation in current AI video generation workflows. Creating one strong clip is becoming easier. Keeping ten or twelve clips visually connected is still much harder.
CANVAS is designed to reduce that gap by treating continuity as a memory problem rather than relying on the model to infer everything again for every shot.
A²RD is where the ten-minute demo becomes relevant
A²RD focuses on extending video over much longer durations.
Instead of trying to create the entire sequence at once, it generates video segment by segment while maintaining multimodal memory of what has already happened.
Google describes a repeated retrieve-synthesize-refine-update loop. The system can extrapolate to move the narrative forward or interpolate when it needs to anchor a new segment to existing characters and environments.
This is the system Google used for its ten-minute demonstration.
For marketers, however, the duration itself is not the main point. Very few businesses need a ten-minute AI-generated film simply because one can be made.
The more useful question is whether the same architecture can make common formats easier to produce: 30 to 60-second paid ads, two-minute product explainers, customer stories, onboarding videos and training content.
If coherent long-form video generation reduces continuity problems across those formats, the business value becomes much clearer. The ten-minute film shows that the architecture can stretch. It does not prove that brands need ten-minute AI videos.
VQQA adds automated critique and prompt refinement
VQQA focuses on quality control.
The system generates visual questions based on what should appear in a video, uses a vision-language model to critique the output, and turns that feedback into prompt changes for the next generation.
Importantly, VQQA is a prompt optimizer rather than a pixel-level editor. It tries to correct high-level problems such as attribute binding or inconsistent character details by improving the instructions used for the next video.
It also uses a global selection mechanism so the system does not automatically accept the latest version. Instead, a rater compares outputs against the original prompt and selects the strongest candidate.
For agentic video generation, that feedback loop is important. Automation without evaluation can simply produce mistakes faster.
What this could change for marketers
There is an important limitation to keep in mind: Google has not launched a finished marketing product here.
There is no new button in Google Ads, Gemini or Google Vids that automatically creates a perfect multi-minute branded film. This is research. But the direction could address several production problems marketing teams already have.
Product explainers could require less cleanup
Longer AI-generated explainers still need manual correction when interfaces, products or presenters drift between scenes. Persistent visual memory could reduce some of that cleanup and make AI video more practical for formats that need more than a few seconds.
Creative variations could become easier to produce
Paid teams rarely need one video. They need multiple hooks, openings, audiences and lengths.
A system that can preserve the same visual world while changing the narrative could make creative testing more manageable. That becomes more relevant as video takes a larger role across paid channels, including developments such as Demand Gen VTC optimization.
Brand consistency could become part of the generation system
CANVAS also points toward a more useful future for branded AI video. Instead of repeatedly explaining how a product or visual identity should look, future systems could potentially remember approved assets, environments, characters and visual references.
That is an inference from the research, not a Google product feature today. But persistent brand memory would be far more useful for teams producing assets at scale than starting every prompt from zero.
Google’s benchmarks are useful, but they are not campaign results
Google created specialized benchmarks for different parts of the research, including GenAD-Bench, HardContinuityBench and LVBench-C.
GenAD-Bench is especially relevant to marketers. Google says it includes 400 scenarios across 50 fictional brands with four products each, and reports a peak Co-Director quality score of 81.4 on the benchmark.
Those results support the technical direction, but they should not be read as proof that all four systems have been validated together as one production-ready marketing pipeline. Each framework targets a different part of the problem and is evaluated separately.
The benchmarks also do not tell marketers whether these videos will improve CTR, reduce CAC or generate more pipeline.
Continuity is a production metric. Campaign performance still needs to be measured separately.
What marketing teams should watch next
The first thing to watch is product availability. The marketing relevance changes quickly if these ideas move from research into Veo, Flow, Gemini, Google Vids or Google Ads.
The second is control. Marketing teams need more than continuity. They need accurate product details, approved claims, brand colors, typography and repeatable creative rules.
The third is cost. Agentic systems can involve several models, repeated generations and evaluation loops. Better output matters, but so does the cost of producing it at campaign scale.
Finally, watch integrations. The larger opportunity appears when systems like these connect with asset libraries, brand systems and campaign workflows rather than operating as standalone research demos.
OneMetrik Takeaway
The important part of Google’s research is not that AI can generate a ten-minute video. It is the architecture required to keep that video coherent.
Google is treating long-form AI video as a systems problem involving planning, memory, generation, evaluation and correction rather than expecting a single prompt or model to handle everything.
For marketers, that is the development worth watching. If these ideas eventually move into products such as Veo, Flow, Gemini or Google Vids, AI video could become more useful for explainers, product demos and campaign creative that need consistency across multiple scenes.
For now, however, this remains research rather than a production-ready marketing workflow.