Katechon ArcThe numbers first
We audited our first five short-drama series in October 2026: 195 episodes, 4.1 hours of runtime, $547 of video. Here is what cost us money and time.
| Problem | Measured cost |
|---|---|
| Shots rendered again | 15% of video spend ($70) |
| A pricier model where a cheaper one would have done | $101 extra |
| Stalls waiting on jobs, retries and fixes | About 50 logged hours |
| Clips rendered but not used (a fast-cut film) | 608 of 972 in BROODFALL |
1. Drift: faces, costumes, props and places
Every clip is a fresh generation, so anything not pinned by a reference wanders. We saw helmets flip on and off between cuts in Dragon Rising, a small character grow wings and night scenes drift to daylight in The Little Fairy Bakery, and style rules leak into places they did not belong: "the red balloon is the only red" painted balloons into eight set sheets in RatKid and the Red Balloon. The fixes were references for every recurring character and location, explicit per-shot state (who wears a helmet, what time of day), and style rules phrased as what a set contains rather than what it must not.
2. Re-renders
Re-renders were the biggest controllable cost. Most came from shots whose descriptions asked for two actions at once, put lighting words on skin, or left the model to guess who was in frame. Writing one action per shot with a concrete light source, approving storyboard frames before video, and regenerating single shots rather than whole episodes kept later series well under the 15% average.
3. Failed and blocked generations
Some jobs simply fail. On Seedance 2.0 Fast about one clip in fifteen failed once and passed on a retry, so retries need to be automatic. A different failure looks the same but is not random: a generated face that resembles a famous actor gets blocked with an error saying the video "may be related to copyright restrictions". In BROODFALL, a lead with a shaved head and grey beard had about 5% of his shots blocked. Rewriting the shots did nothing; changing the character's face fixed it. If one character's shots keep failing, recast the face and re-render all of that character's shots so he stays consistent.
4. Running out of provider credit mid-production
Twice a provider balance ran out partway through a film. RatKid and the Red Balloon switched video providers mid-production and still held its cast because the references lived in the project, not in the provider. BROODFALL stopped after 972 clips and became a 10-minute cut built only from what existed. Price the whole series before you start, and check the balance covers the batch before you queue it.
5. Sound
Native-audio models occasionally invent a voice-over or a stray line on near-empty shots. Raw renders also came out quiet, around -17 to -19 LUFS, and several early series shipped with no music at all. The later ones were mixed to about -14 LUFS with music placed in the edit. Listen to every clip before captioning.
6. Captions
Captions timed from the script sat a line early; aligning the script to speech-recognition word timings fixed 396 of 400 lines on Blood Sun: The Fall of Kratos. Speech recognition, in turn, misspelled invented names: six errors in the first caption track of The Fall of Kalak.
7. Too much complexity
Our most complex series had the most drift, the most re-renders and the most review time. The simplest one, High Note, had about 5 core characters and 6–8 sets, rendered every episode on one budget model, and was the easiest to keep consistent. Complexity multiplies every problem above.
8. Changing things after the fact
Two late changes were cheap because of how the projects were built. Renaming Blood Sun's cast after production touched only spoken lines, captions and titles, so 25 shots were re-rendered instead of the series. Switching the video model mid-project was not cheap: the provider for a queued shot is chosen when it is dispatched, so changing the project's model also changed shots already in the queue.
A pre-flight checklist
- Price the series and confirm the budget covers it with 15% for retakes.
- Keep the cast and set count small; make a reference for each.
- Write per-shot state for anything that can flip: costume, helmet, time of day.
- One action and one light source per shot; colour words on the environment.
- Approve storyboard frames before video.
- Pick one video model for the whole series and do not switch mid-queue.
- Retry failures automatically; investigate a character whose shots keep failing.
- Mix sound and align captions after the edit, then check names.
Where Katechon Arc fits
Arc was shaped by these failures: references attached to every shot, storyboards before video, a credit estimate and balance check before each batch, automatic retries and single-shot regeneration.
Arc cannot stop a provider from blocking a lookalike face or guarantee a take is right first time, and sound mixing and captions still happen outside it.
People also ask
How many AI shots have to be re-rendered?
In our first five series, re-renders took 15% of video spend. Budget about 15% extra for a planned series and more for an edit-heavy film.
Why do AI characters change between shots?
Each clip is generated fresh. Without a reference image attached to every shot, the model reinterprets the character's description each time.