Quick answer: the pipeline is plan, quote, approve, render, place dialogue, deliver, and only the fourth step is the video model. Dialogue placement and padding are deterministic file operations that cost no generation credits, and timed text is a deliverable rather than an export option.
Every "AI film" demo collapses into the same shape: a prompt box, a few clips, a montage. Making an episode is a different problem, because an episode has structure — scenes, shots, dialogue that has to line up, a budget somebody approved, and deliverables that a player or an editor can ingest.
H3 is the shot renderer in that pipeline on Callirra. This post is about the pipeline, because that is the part people underestimate.
The unit of work is a shot, not a prompt
A shot in production terms has: a script line, a duration, a framing, a subject, and a relationship to the shot before and after it. When that becomes a generation job, four of those five are parameters, and the fifth is the reason a shot list exists at all.
Practically, the pipeline is:
- Plan — the story is broken into episodes, and episodes into shots, with durations.
- Quote — each shot is priced from its intended tier and duration before anything renders.
- Approve — a budget covers the plan; the run stays inside it.
- Render — shots go to the video model, H3 among them.
- Place dialogue — voice tracks are positioned at a start time and padded to the shot's length.
- Deliver — a master file plus subtitles and a machine-readable manifest.
Steps 2 and 3 are the ones that separate a pipeline from a hobby. A plan that cannot be quoted cannot be approved, and a run that cannot be stopped mid-way has no budget at all.
The job a video model cannot do: timing audio to picture
Dialogue is where most generated episodes fall apart. Three lines at 3.55, 3.55 and 3.175 seconds, placed at 0.5, 5.5 and 10.5 seconds on a fifteen-second timeline — that is a real mix from one of our acceptance runs. The video model did not produce that arrangement; the pipeline did.
Two primitives make it work:
- Placement. A voice clip has an intended entry point, not just a duration. Without an explicit start time, every line lands at zero and the scene becomes a pile-up.
- Padding. A 3.175-second line on a six-second shot needs 2.825 seconds of silence after it, or the shot gets cut to the shortest element. Padding is a deterministic file operation — no generation credits — and it is what lets a short line live on a long shot.
If you take one thing from this post: generate the picture to the length of the performance, not the performance to the length of the picture.
Subtitles are a deliverable, not an export option
Timed text is where "we made a film" turns into "we made something a platform will accept". The pipeline produces:
| Deliverable | What it is |
|---|---|
| MP4 | the finished mix, saved as a private asset |
| SRT | subtitle file in the project language |
| VTT | the same timing, for players that want WebVTT |
| Render JSON | an editable manifest of the render, not an interchange format |
The last row is the one to be careful about. That JSON describes the render; it is not FCPXML, OTIO or EDL, and it does not promise a lossless round trip into a desktop NLE. If your finishing house needs an NLE project, plan on rebuilding the timeline, not importing the manifest.
What this costs, roughly
Per the published H3 rates, ten shots of five seconds at 768P is 100 credits against the shot budget; ten shots of ten seconds at 1080P is 300. Dialogue lines and subtitle generation are separate, much smaller line items — in the acceptance run quoted above, the whole localisation pass was 0.67 credits.
The budget-shaped insight: picture is the expensive half, sound and text are nearly free. Plan the money into the shots.
FAQ
Can the video model time dialogue to picture? No. Placement and padding in the pipeline do that, and both are deterministic file operations.
What is the render JSON for? Logging and editability of the render. It is not FCPXML, OTIO or EDL and does not round-trip into a desktop NLE.
Where does the budget actually go? Picture. Ten five-second 768P shots are 100 credits, while the whole localisation pass in our acceptance run was 0.67.
Where this runs
The episode workflow — plan, quote, approve, render, place dialogue, deliver — is the AI Filmmaker workflow on Callirra (https://callirra.com/stories), with H3 available as a shot model. New authoring requires an active Starter subscription or higher; work you already own stays readable, downloadable and exportable after a term ends.
I work on Callirra, so the mix timings and deliverables above come from our acceptance records rather than a demo script. If you build a similar pipeline on another provider, the three primitives — placement, padding, timed text — are the parts worth copying.