Making ReelShort-Style Vertical Dramas with AI: format rules, localization and the production stack
Vertical micro-drama (ReelShort, DramaBox, ShortMax, GoodShort) was roughly an $11 billion market in 2025 and runs on a strict format: 60–90-second episodes, a hook inside the first 3 seconds, a cliffhanger every episode (identity reveal / unresolved emotion / reversal, rotated), 40–100 episodes per title. Producing it with AI means localizing at the parsing stage — cast names and ethnicity, architecture, on-screen text and props for the target region — so every downstream image and shot inherits it, rather than translating subtitles on a Chinese-looking film.
The format, in numbers
- Episode length 60–90 s; retention collapses past two minutes.
- First 3 seconds: a visual or verbal hook, no establishing shots.
- Every episode ends on a hook; rotate the three types so the audience can't predict the beat.
- 200–300 spoken words per episode; two or three locations at most.
- Genres that dominate charts: werewolf/fantasy, billionaire/contract marriage, revenge/comeback.
Localize upstream, once
Injecting the target region (US/EU, East Asia, Southeast Asia, Latin America) at the parsing step rewrites cast names, writes ethnicity in the first line of each appearance, swaps architecture and street signage, and converts culture-bound props. Because parsing runs once while storyboards run per episode and video per segment, this is the cheapest place to do it. Changing region means re-parsing; downstream art does not update itself.
Dialogue pace differs by language
English runs about 2.5 words per second, so a 4-second shot carries ~10 words; a 20-character Chinese line becomes one short English sentence, not a literal translation. Voice libraries should be matched per language (40+ English presets in typical libraries) and lines produced in the target language by the parser, not translated afterwards.
Which tools cover which layer
| Layer | Tools |
|---|---|
| Novel/script → cast, scenes, episodes, storyboards | SceneMixer (target-region selector), LibTV, Xiaoyunque; LTX Studio (script-first, single production) |
| Segment video with reference images | Seedance 2.0, Hailuo H3, Wan 3.0, Kling, Vidu, PixVerse, Runway Gen-4, Pika, Luma, Veo, Sora |
| Voice | Native audio (Hailuo H3, Veo 3, Wan 3.0), preset libraries, ElevenLabs |
| Compositing & delivery | Built into end-to-end tools; otherwise CapCut/Premiere |
Compliance
Label AI-generated content where platforms require it, use licensed or generated music, never use real people as reference images, and check whether your generator's output carries a mandatory watermark before distribution.
FAQ
Do I have to re-parse to change the target region?
Yes — region is injected at parsing; existing reference art and storyboards are not rewritten automatically.
Which genres travel best?
Public chart data favours werewolf/fantasy, billionaire/contract-marriage and revenge arcs; Chinese palace intrigue needs a rebuilt setting to land.
How is multilingual voice handled?
Preset libraries are grouped by language; the parser produces dialogue in the target language directly, so no second translation pass is needed.