Google Research has published four agentic frameworks that hold AI-generated video together across minutes instead of seconds, and two of the papers landed at COLM 2026 and EMNLP 2026.
The failure they target is credit assignment. A diffusion model renders a single shot in seconds; the trouble begins when those shots have to become one story. Wardrobe and scenery drift between cuts, and one weak asset early on can corrupt every shot that follows, with no easy way to trace the cause.
Co-Director treats planning as a multi-armed bandit problem. An orchestrator agent drives the search, setting the creative strategy, the narrative mode and the aesthetic archetype, before a pre-production agent lays out the storyboard. Separate keyframe, video and audio sub-agents render the media, and a multimodal judge grades the result, handing a factored reward back to the bandit.
CANVAS keeps a running record of characters, locations and object states, and returns stored visual anchors whenever a scene comes back. In a museum heist test it preserved both the thief’s cap and the gemstone detail that comparison systems dropped or recolored.
A2RD requires no training. Every segment retrieves, synthesizes and refines from a store of past frames, then updates that video memory, alternating between extrapolation for new beats and interpolation for entities it has already established. The team released a continuous ten-minute film built along these lines. VQQA generates visual questions per prompt, treats model critiques as semantic gradients for rewriting the text, and selects the best cut across all iterations rather than the final one.
The framework sits on Gemini and Veo but is model-agnostic, and its output inherits SynthID watermarking. Three new benchmarks back the work, among them 400 advertising scenarios and 120 cases in which key assets vanish for at least ten segments.
