Google Research presented a multi-agent framework designed to generate coherent, minutes-long video narratives on top of Gemini and Veo. The system addresses character drift, changing environments, and cascading errors that occur in linear generation pipelines. Its components plan visual continuity, maintain persistent storyboard state, generate scenes sequentially, and refine prompts from automated visual critiques. Google reported stronger consistency in its evaluations, while outputs inherit SynthID watermarking from the underlying models.
Source: https://research.google/blog/coherent-long-form-video-generation/
