Google's co-director wants to end character drift in AI video

Google Research published a way to stop AI video from losing its characters halfway through. On September 24, 2026, researchers Yale Song and Yiwen Song described a multi-agent framework they call an AI video co-director, built as an orchestration layer on top of Gemini and Veo, which plans continuity across shots instead of chaining one prompt after another.

The problem it targets is the one every creator hits by the third shot. A pipeline that generates clips one at a time has no memory of the last one, so a jacket changes colour, a face drifts a few years, and a background quietly rearranges itself. Google's blog post calls these semantic drift and feature drift, and notes the compounding version: one bad asset upstream corrupts everything generated after it, which the team describes as a cascading failure and a credit assignment problem, because the final bad frame is hard to trace back to the prompt that caused it.

What the framework actually does

The paper formalises video storytelling as a global optimisation problem. An Orchestrator Agent picks a creative configuration across three axes, which the team names creative strategy, narrative mode and aesthetic archetype, by treating the choice as a multi-armed bandit problem. In plain terms, the system is searching a space of ways to tell the story while it also decides what the story looks like.

From there the work splits. A Pre-Production Agent turns the chosen configuration into a scene-by-scene storyline and a storyboard. A Production Agent renders it with sub-agents that have one job each: a keyframe agent anchors characters and scenes, a video agent adds motion, and an audio agent layers voiceover and score. Then a multimodal model acts as a judge, scoring the assembled cut against the three original axes and feeding that score back so the search can do better on the next pass.

That feedback loop is the interesting design decision. The system treats quality as something measured at the end and optimised backwards, rather than something assumed from a good prompt.

The four pieces

The blog presents the work as four frameworks rather than one system.

Co-Director, accepted at COLM 2026, is the orchestration layer described above. CANVAS, accepted at EMNLP 2026, handles visual storyboarding with what the team calls structured visual memory, aimed at keeping a scene looking like itself when it reappears. A²RD attacks long-duration temporal stability, which is where layout drift shows up across continuous minutes. VQQA closes the loop on prompt quality, rewriting prompts to fix physical and compositional errors it finds in the output.

Each has its own paper, and the blog links all four. Reported results include a peak quality score of 81.4 on GenAD-Bench for the co-director, continuity gains on ST-Bench and HardContinuityBench for CANVAS, layout drift reductions on VBench-Long and LVBench-C for A²RD, and quality gains on T2V-CompBench, VBench2 and VBench-I2V for VQQA. Those are the team's own evaluations, run against benchmarks the team also links, and none of it is an independent third-party test of the system as a whole.

What it is not

There is no product here. Google has published no waitlist, no API, no pricing and no access date for the co-director, and the blog frames the work as research with human-in-the-loop workflows listed as a future direction. The researchers are explicit that the goal is to take over the tedious parts, "ensuring they remain the control of creative direction and narrative design," which is a description of an assistant, not a replacement.

One practical consequence is worth noting for anyone generating commercially. Because the framework sits on Gemini and Veo, everything it produces inherits SynthID, Google's invisible watermark, and the blog says safety classifiers can also be applied across the finished video to catch problems that arise from combining clips that were each fine on their own. That is a provenance answer built into the pipeline rather than bolted on afterwards. Google's own SynthID page notes that ordinary users can check a file by uploading it to Gemini and asking whether Google AI made or altered it, and that the company runs a separate SynthID Detector portal for journalists and media professionals.

For creators the honest reading is this: the fix for character drift is being researched at the level of planning and memory, not at the level of a better prompt, and the layer that solves it will probably sit above whichever video model you already pay for. Google says the architecture is model-agnostic, which is the part that travels. Whether the promised control arrives as a feature inside a Google product or as a pattern other tools copy is the open question the blog does not answer.

Try the diagnostic yourself this week: generate six shots of the same character from the same reference image with whatever tool you use, put the frames side by side, and count how many keep the face. That count is the problem these papers are trying to solve.