Google Research has unveiled Co-Director, an agentic framework designed to generate coherent, multi-minute videos. The system addresses common issues in video generation, such as semantic drift and cascading failures, by treating video production as a credit assignment challenge. It operates on top of Gemini and Veo models and is model-agnostic, allowing integration with other generators while utilizing SynthID watermarking.
The framework comprises four distinct approaches. Co-Director utilizes a multi-armed bandit and an Orchestrator Agent to manage creative strategy and narrative flow. CANVAS maintains consistency of characters and objects across scenes by tracking states and retrieving visual anchors. A2RD is a training-free architecture that uses a memory loop to handle story beats, recently demonstrated in a 10-minute film.
The fourth component, VQQA, employs visual question answering and VLM critiques to refine prompts without needing access to model internals. To evaluate these systems, Google introduced three new benchmarks: GenAD-Bench, which includes 400 ad scenarios; HardContinuityBench, focusing on prop and scene consistency; and LVBench-C, testing the persistence of assets over long sequences.
The research behind these frameworks has been accepted at major conferences, including COLM 2026 and EMNLP 2026. The system architecture allows for iterative improvements, where a global selection process identifies the highest-quality video output across multiple iterations, rather than relying solely on the final generated result.
