SAGE: Self-Evolving Storyboard Skills via Attribution-Guided Rule Evolution

Details

Storyboarding — converting screenplays into visual shot plans — requires tacit directorial expertise that is hard to capture in rules or prompts. SAGE (Skill with Attribution-Guided Evolution) addresses this through a deployed framework that learns directing rules from expert demonstrations and continuously refines them. The workflow has four stages: SAGE derives content-independent rules by contrasting each training screenplay with its expert storyboard; during generation, it records which rules were applied to each narrative group; localized feedback enables targeted updates to individual rules rather than bulk retraining; evolved rules are packaged by scenario with a routing index, so each narrative group retrieves only contextually appropriate rules. Production deployment on Virtual Film Studio over 14 days processed 1,344 narrative group outputs. 87.2% were accepted without substantive edits. The production team recorded 83% less authoring time per episode. On 18 test episodes across three genres, SAGE scored 77.8 vs 77.1 for professional directors on a rubric validated by experts. The authors also release PROSE, the first public dataset pairing screenplays with professional storyboards across 68 episodes. The deployment numbers — not just benchmark scores — make this particularly credible as a real-world AI content creation success story.