Guides
Budgeting AI Shots: Batch Your Notes, Test for a Tenfold Price Rise and Lock Tool Versions
ยท RenderBob team
Batch review notes, lock preproduction, test a tenfold price rise and pin tool versions. A guide from the Korean producers' 8 November cost warning.

The Korean producers' warning on 8 November points to a practical lesson: in AI production, the number of revision rounds and the price per attempt drive the budget more than the headline cost. This guide turns their points into steps.
Count attempts, not shots
Our 3 November guide showed that cost per usable clip depends on how many attempts you discard. Kim Kwang-jib's point adds the human side: every re-render is paid for again. Use his example as a rule. Ten notes sent together can cost one re-render, while ten notes sent across ten days can cost ten.
- Collect notes into one review round per sequence.
- Set a limit on rounds before the shot is locked.
- Record each re-render and its reason.
Lock the preproduction
Kim says weak preproduction can push AI above the 35mm era in cost. Before generating, fix the script, storyboard, character references and style references for a sequence. Treat changing them as a new budget line.
Test the budget against a price rise
The tenfold stress test is a planning step we added. Take your current price per second, multiply it by ten, and see whether the project still works. Kim's own example turned a token bill of roughly KRW500 million into KRW5 billion, which equals the whole KRW3-5 billion budget of a typical Asian feature, while a far larger Hollywood budget could absorb it. If your plan fails at ten times, either budget a reserve or choose a cheaper path for those shots.
Lock tool versions per sequence
Lee says tools update weekly and make a year-old shot look dated. Mixing versions inside one scene can break continuity. Choose a model and version per sequence, note it in the prompt log, and change it only at a sequence boundary unless a shot fails.
Use AI where it is cheapest to be wrong
Kim's hybrid approach puts AI on large crowds and live action on leads' close-ups. Generate wide, forgiving shots, film the shots that audiences judge on faces, and review the split after a test sequence.
Test delivery early
Lee says streamers check files mechanically and AI output struggles with bitrate and colour grading. Export a short test clip in the delivery format at the start.
Keep the cheap shots cheap
Match the model to the shot: a lower-cost model for establishing shots, a stronger one for hero shots. Compare cost per usable clip, as in our 3 November piece.
What to watch
Whether tool makers publish stable price commitments, and whether more productions report their revision counts alongside their budgets.
RenderBob 3 and this story
RenderBob 3 makes the model a per-clip choice and keeps each clip's references and prompt log with the studio, which is what a revision count, a version lock and a cost comparison depend on.
More from the blog
- AI-Made Political Ads: The Counts Keep Growing, and Most Carry No AI Label. What Creators and Agencies Should Take From It
Wesleyan Media Project counts of AI-made US political ads rose from 33 in May to 220 by 11 October. A 3 November report cites nearly 300. Most of the counted ads carried no AI disclaimer.
- Price per Second Is Not Cost per Usable Clip: A Practical Test for Cheap Video APIs
A video API at half a cent a second can still cost more per finished shot than a dearer one, if more of its clips are thrown away. Measure cost per usable clip on your own images.