LLM Shot Planner

Key Insight

A minute-long story is too much for any single text-to-video (T2V) generation, so this project borrows the film world's solution: a shot list. A small large language model (LLM) acts as a "director," expanding a one-line prompt like "a knight rescues a princess" into a structured JSON plan — a sequence of shots, each with its own description — that you then generate one at a time and stitch together. This is the planning half of hierarchical generation: the LLM decides what happens in what order before any pixels are made, which is the only way to get a video whose events follow a sensible narrative rather than wandering. You evaluate the result on coherence — do the shots actually tell the intended story, and does the knight stay the same knight across them?