An AI video production workflow that locks in key assets first
The lesson from making a 1-minute cinematic video about Yi Sun-sin's Battle of Myeongnyang is the order of steps: the prompt does not go in right away. First, the AI is made to research the topic, the Battle of Myeongnyang, and how to use the generation tool, separately, to fill the context; the two results are then combined and metaprompting is used to produce a tool-specific prompt containing a second-by-second shot plan. In the generation stage, key assets such as Admiral Yi Sun-sin, the panokseon warships, the Myeongnyang Strait and the Japanese warships are first created as 2K images with Nano Banana 2 and locked in; eight shots are generated in parallel from them, the images are converted to video, and background music is added to finish the edit. Physics such as cannon fire still looks awkward, but the result comes out for 1 credit or less in about 30 minutes. It is a demo of one specific commercial tool, but the procedure for securing consistency itself — research → metaprompting → lock in key assets → generate shots in parallel — can be carried over to any AI video pipeline.
Key points
- Research the topic and the tool's usage separately, then combine them and metaprompt a tool-specific prompt
- Locking key assets in as 2K images first keeps the shots consistent with one another
- Generate eight shots in parallel from a second-by-second plan, then convert image → video
- Physics looks awkward, but a 1-minute clip takes 1 credit or less and about 30 minutes
This is a note: my own write-up after reading several sources. It has no source link; events checked against a source are kept as records.