Why short drama suits AI
Short drama is structurally the format AI video handles best, for reasons that have nothing to do with model quality.
- Short shots. Episodes are built from many brief beats, and AI holds four to six seconds far better than thirty.
- Vertical framing, which hides composition weaknesses and matches consumption.
- Dialogue-driven coverage — two-shot and close-up patterns that do not demand complex camera work.
- High audience tolerance for stylisation in genre content.
The fit is structural: these are format properties, not model properties. Short drama was always going to be the first genre AI production cracked.
What it actually costs
Reported cost reductions are large — per-minute production cost falling by roughly three quarters in some markets, with full short-form projects achievable at a fraction of traditional budgets. Treat the direction as reliable and the specific numbers as market-dependent.
| Cost line | Traditional | AI-assisted |
|---|---|---|
| Per minute production | High, crew and location dependent | Reported down by roughly 70-75% in some markets |
| Per project | Tens of thousands and up | Can reach low five figures for short form |
| Iteration cost | Reshoots are expensive | Regeneration is cheap |
| Distribution cost | Unchanged | Unchanged — and now the dominant line |
The last row is the one that matters. Cutting production cost does not cut the cost of getting watched, and in a market flooded with supply that cost rises.
The supply explosion and what it means
Thousands of new titles per day are now reported on major platforms, with tens of thousands of production companies having shipped at least one title. That is not a market; it is a lottery with many tickets.
- Marginal production cost approaching zero means supply is unbounded.
- Platform algorithms allocate attention, so they capture most of the value.
- Free ad-supported models dominate in some markets, which means revenue follows watch time rather than quality.
Plan for the supply curve: when anyone can make it, the differentiator is not making it well. It is getting the first fifteen seconds right and publishing consistently.
An episode workflow
-
Script in beats, not scenes
Write each episode as fifteen to twenty-five short beats of four to six seconds. Each beat is one generation.
-
Lock character references first
One canonical reference per character, front-facing and neutral. Everything downstream depends on this.
-
Build a shot list with camera intent
Decide static, push or pan per beat before generating. Deciding after wastes clips.
-
Generate beat by beat, chaining approved frames
Use the approved frame from each beat as the input for the next.
-
Voice separately
Generate clean dialogue audio, then lipsync or cut to coverage. Do not rely on in-generation audio for drama.
-
Cut hard
Short drama is cut fast. Trim any beat that does not advance the hook.
Where consistency breaks
The same failure points appear in nearly every AI drama project.
| Failure | Why it happens | Mitigation |
|---|---|---|
| Face changes between shots | No persistent identity between generations | Chain approved frames; never generate a new shot from text alone |
| Outfit changes mid-scene | Paraphrased character wording | Paste the identical character block every time |
| Lighting jumps | Each generation invents its own lighting | Specify lighting per shot and grade to match |
| Voice differs between episodes | Voice settings drift or are not saved | Lock and save a voice profile; regenerate all dialogue from it |
| Background discontinuity | Locations reconstructed per shot | Use one location reference image per set |
Why hooks matter more than quality
- In a free, ad-supported model, revenue follows watch time, and watch time is decided in the first seconds.
- Algorithmic distribution means a weak hook never gets a chance to demonstrate quality later.
- Publishing cadence compounds: more titles means more chances for one to land.
Budget your effort accordingly. Polish on episode forty matters less than the hook on episode one.
Does the money work
-
Model revenue per thousand views, not per title
Platform revenue sharing varies widely and changes often. Get current terms in writing.
-
Count your true cost per episode
Generation plus editing plus voice plus your time. Most projects undercount time badly.
-
Assume most titles earn nothing
With unbounded supply, returns are heavily skewed. Budget for a hit rate, not an average.
-
Check disclosure requirements
Several jurisdictions now require labelling of AI-generated content. Non-compliance risks removal after you have spent the production cost.
The honest summary: AI made short drama cheap to make and did nothing to make it easy to monetise. Enter with a distribution plan, not just a production capability.