OmniScript: Towards Audio-Visual Script Generation for Long-Form Cinematic Video

OmniScript addresses a critical gap in Multimodal LLM (MLLM) capabilities: the transition from short-clip captioning to long-form cinematic understanding. While the project is very new (0 stars, 4 forks, 4 days old), its value lies in the 'first-of-its-kind' human-annotated dataset and the formalization of the Video-to-Script (V2S) task. Defensibility is currently low (4) because the project is a research artifact rather than a product with a network effect. The 'moat' is essentially the dataset, which is expensive to replicate but once published, serves as a benchmark for larger players. The frontier risk is 'high' because Google (Gemini 1.5 Pro) and OpenAI (GPT-4o) are aggressively expanding long-context video windows (1M+ tokens). Gemini 1.5 Pro already demonstrates zero-shot capabilities in video understanding that threaten specialized research models. Platform domination risk is high as this functionality is a natural extension for Adobe (Premiere Pro/Frame.io) or OpenAI (Sora/Editor tools). A displacement horizon of 6 months is estimated because the underlying MLLM architectures (like LLaVA or Qwen-VL) used in such research are being rapidly superseded by frontier lab releases that handle long-form video natively without the need for the specific hierarchical engineering proposed here.