Researchers introduce a unified multi-agent orchestration layer designed to produce coherent, multi-shot videos that preserve visual continuity over minutes-long narratives. The system sits on top of foundation models Gemini and Veo, inheriting native safety features such as SynthID watermarking, and reframes long-form generation as a global optimization and world-state tracking problem.
The work is organized into four complementary frameworks: Co-Director (a hierarchical AI co-director), CANVAS (continuity-aware visual storyboarding), A²RD (segmental autoregressive video synthesis), and VQQA (closed-loop video quality question answering). Together these tools coordinate creative intent, maintain persistent visual memory, scale temporal dynamics, and perform iterative test-time prompt refinement.
Motivation
Although modern diffusion-based models can produce high-fidelity short clips in seconds, they struggle to maintain semantic and visual consistency across long, multi-shot sequences. Common failure modes include semantic drift (subtle changes in clothing or scenery), cascading failures (early artifacts corrupting downstream synthesis), feature drift, and content collapse where narratives cease to advance meaningfully. These issues are compounded by the difficulty of attributing terminal failures to specific prompts or pipeline stages.
Co-Director: hierarchical global optimization
Co-Director formalizes storytelling as a hierarchical multi-agent global optimization. Instead of linear prompt chains, it uses hierarchical parameterization and a multi-armed bandit (MAB) algorithm to select creative configurations across three dimensions: Creative Strategy (intent), Narrative Mode (story structure), and Aesthetic Archetype (visual tone and cinematography). The MAB balances exploration of new narrative strategies and exploitation of effective configurations.
The pipeline runs two interconnected loops: strategic steering and multi-stage production. The Orchestrator Agent selects a creative configuration via the MAB. The Pre-Production Agent synthesizes a scene-by-scene storyboard and visual assets. The Production Agent executes the storyboard using specialized sub-agents: a Keyframe Agent for anchor visuals, a Video Agent to add motion, and an Audio Agent for voiceover and score. A multimodal LLM Judge critiques the compiled cut along the three designated dimensions and returns a factored reward to the MAB to iteratively refine subsequent generations.
Because Co-Director operates as an orchestration layer over Gemini and Veo, it feeds structured prompts into those foundation models; the architecture is model-agnostic and can work atop other generative foundations. Generated media inherit native protections like SynthID, and additional safety classifiers can be applied across final outputs.
CANVAS: continuity-aware visual storyboarding
CANVAS (Continuity-Aware Narratives via Visual Agentic Storyboarding) enforces visual coherence by maintaining structured representations of characters, locations, and object states as the narrative evolves. Built as an orchestration layer over Gemini, CANVAS uses a persistent visual memory to retrieve anchors or initialize new ones, ensuring smooth transitions and preserving identity and spatial structure when scenes are revisited.
In comparative examples, Gemini-3.1-Pro used alone exhibited prop inconsistency and background drift across cuts. An alternate multi-agent framework, AutoStudio, likewise showed degradation such as character drift. CANVAS’s persistent visual memory produced more consistent results for recurring characters, spatial geometry, and object states across multi-shot sequences.
A²RD: scaling temporal dynamics
A²RD is an agentic autoregressive architecture for translating storyboards into minutes-long video. It generates segment-by-segment and maintains a multimodal video memory that tracks contexts and dynamics. Each segment follows a retrieve–synthesize–refine–update loop, and the agent adaptively chooses between extrapolation (to push the plot forward) and interpolation (to anchor returning entities and environments).
This adaptive switching balances narrative progression with physical coherence, preserving character identity, costume details, and scene geometry across long durations. The authors demonstrate the approach on a ten-minute generated movie to stress long-horizon consistency.
VQQA: closed-loop refinement
VQQA (Video Quality Question Answering) provides an efficient, black-box method for identifying and correcting visual artifacts without requiring white-box access or heavy computation. It dynamically generates visual questions tailored to the prompt, and uses Vision-Language Model (VLM) critiques as semantic gradients—natural-language directional feedback—to iteratively refine prompts.
Instead of pixel-level editing, VQQA refines text prompts to resolve compositional defects such as attribute binding errors or inconsistent character attributes. To avoid semantic drift during iterative refinement, a Global Selection mechanism has a global VLM rater evaluate every video produced along the optimization trajectory against the original prompt and select the highest-scoring candidate. Practical examples include mapping a realistic mylar balloon texture onto a strict cuboid geometry and preserving instrument assignments for musicians across cuts.
Benchmarks and key results
The team developed three benchmarks to evaluate challenges relevant to professional video production:
- GenAD-Bench: a human-in-the-loop pipeline pairing Gemini 3 Pro with specialized image models to generate 50 fictional brands with four distinct products each, yielding 400 unique scenarios to test exact marketing constraints.
- HardContinuityBench: focuses on spatial and environmental continuity using detailed GPT-5.2 multi-shot storyboards, with large gaps between reappearances and frequent costume/accessory changes.
- LVBench-C: examines long-horizon temporal dynamics across 120 text-only scenarios that enforce a strict gap rule (critical visual assets must disappear for at least 10 segments before reappearing with narrative-driven changes).
Across evaluations, Co-Director reached a peak quality score of 81.4 on GenAD-Bench and improved story consistency on ViStoryBench. CANVAS reduced scene drift across reappearances on ST-Bench and HardContinuityBench. A²RD minimized layout drift and improved character and environment consistency in continuous multi-minute runs on VBench-Long and LVBench-C. VQQA’s closed-loop prompt optimization delivered measurable absolute quality gains on T2V-CompBench, VBench2, and VBench-I2V. Detailed model architectures, training configurations, and baseline comparisons are provided in the individual papers.
Future work
The authors position these frameworks as foundational steps toward enabling coherent long-horizon visual storytelling for creators. Ongoing work aims to incorporate human-in-the-loop workflows so that creators retain narrative control while the system handles temporal consistency and world-state tracking.
Acknowledgments
The research was supported by broader teams at Google. The authors acknowledge contributions from Andrew Pan, Brett Slatkin, Burak Gokturk, Carina Claassen, Daniel Vlasic, Do Xuan Long, Ishani Mondal, Jasmine Leon, Jingyun Liu, Joe Timmons, Jordan Boyd-Graber, Khanh G. LeViet, Kuang Su, Long T Le, Mihir Parmar, Min-Yen Kan, Nathan Hodson, Nick Losier, Palash Goyal, Rhyard Zhu, Sebastian Ko, Scott Penberthy, Yan Xu, Yang Li, Ye Jin, and Tomas Pfister.
