Google Research has described Co-Director, a research system that coordinates language, image, video and audio models to make short ads. After each run, a multimodal AI judge scores the finished cut and the creative choices behind it; Co-Director keeps the best-rated video and uses the direction ratings to guide what to try next, aiming to keep characters, products and settings consistent across shots.
Many video pipelines pass each model's output straight to the next. A flawed storyboard can then carry into later clips. Co-Director adds a planning layer: one agent chooses a broad creative approach, another drafts the story and storyboard, and production agents create keyframes, video and sound.
A multimodal judge gives each finished cut an overall score for temporal coherence, visual quality and consistency; that score helps select the best video. It separately scores the creative strategy, story format and visual style. Those direction scores update a multi-armed bandit, a search method that tries several options and increasingly favors those associated with stronger feedback. Agents can also ask models to critique and regenerate weak storylines or keyframes rather than accept every intermediate output.
The paper tests Co-Director on GenAd-Bench, 400 scenarios built around fictional products. In the reported setup, each output is a 12-second video with four shots; Google models generate the assets, clips and audio, then Gemini 3 Pro scores the result. That test probes whether a pipeline can follow constrained creative directions, but it does not establish how Co-Director performs with real brands or creators.
The researchers also had five people rate each video in a 50-scenario subset. The paper reports that its AI judge tracked human ratings less closely for visual quality than for audience alignment or marketing appeal, so its automated score is not a direct substitute for audience judgment.
Google introduced Co-Director as research on Sept. 24; the paper's first arXiv version was posted on April 27. The preprint also evaluates the system on a separate cinematic-narrative benchmark. Google Research presents Co-Director as a research framework, not a released product.