CineForge: Self-Improving Agents for Long-Horizon Video Generation
CineForge turns past production failures into stage-specific policy updates for future long-form video runs.
The paper describes a video-production agent that decomposes stories, tracks narrative and visual state, designs shots, builds prompts, renders clips, and records the process as a production trajectory. Its evolution module reviews those trajectories, turns recurring issues into bounded patches, and validates them through replay and paired evaluation. The authors also introduce CineScope, a 100-script benchmark and metric for judging full story realization. On their tests, the evolved policy raises CineScope-Metric from 4.024 to 4.380 and cuts review LLM calls by 37.0% on new stories. ArXiv · AI/CL/LG's note
The paper describes a video-production agent that decomposes stories, tracks narrative and visual state, designs shots, builds prompts, renders clips, and records the process as a production trajectory. Its evolution module reviews those trajectories, turns recurring issues into bounded patches, and validates them through replay and paired evaluation. The authors also introduce CineScope, a 100-script benchmark and metric for judging full story realization. On their tests, the evolved policy raises CineScope-Metric from 4.024 to 4.380 and cuts review LLM calls by 37.0% on new stories. ArXiv · AI/CL/LG's note
score 5