Bring it as it is: import a script (ChatGPT, Word, Final Draft, Fountain), split into episodes, not rewritten; or turn a novel into a series with a season plan you can change. Reading and splitting are free.
The questions creators ask first.
How a vertical drama script is laid out, how long episodes run, and how a season is shaped.
A vertical drama script is a screenplay cut into one-to-two-minute episodes, each with a cold open, three to five short scenes, and a last line that is a question. Write it in standard screenplay blocks, one shot per paragraph, dialogue under twelve words, and put the episode's hook and cliffhanger on their own labelled lines so nobody has to hunt for them.
The working standard for a paid vertical drama is 50 to 80 episodes of one to two minutes, with the first 5 to 12 free. Shorter seasons of 10 to 15 episodes exist on real platforms and suit a self-contained story or a first release. Plan the season in total minutes first, then decide how to cut it.
Treat fifty episodes as five acts of ten, each act with its own promise and its own payoff, and put the season's midpoint reversal between episodes 20 and 30. Give every episode one turn, every act one betrayal or revelation, and never let the lead's goal change; only the means and the allies do.
Don't map chapters to episodes. Secure the adaptation rights first, cut the book down to its engine (one lead, one want, one obstacle, one secret), move the book's best reveal into the first thirty seconds, and rebuild the story as a ladder of one-minute turns where every thought the narrator had becomes a line someone says out loud. Most of the book's subplots, and almost all of its first act, will not survive, and that is the adaptation working.
Don't tell it in order. Decide first what stays true (the feeling, the choice, what it cost you), what you will dramatize (events, timing, who said what) and who gets renamed (everyone but you, unless they agree). Then cut what happened down to one engine: you, the one thing you wanted, the person or force in the way. Open on your worst day, not your childhood, and give every episode one turn. The hardest part is getting the inner story out of your head, which is why it helps to be interviewed rather than to face a blank page.
Start from what the picture shows and the track lacks. Keep every sound that is already there, add effects only for actions a viewer can see or should hear just off screen, lay one ambience bed per scene and a music bed that drops under the dialogue. Keep effects out of the lines unless they are clear impacts, make low hits audible on a phone, and check each effect where it sits in the cut instead of trusting a list.
Seedance, Kling and Veo render each clip with its own audio, and that audio includes music written for that clip alone. Stitch eight clips together and you get eight songs that stop and restart at every cut, in moods that do not connect. The fix is to remove the music from each clip while keeping the dialogue and effects (stem separation), then score the whole timeline once in your editor, change the music only where the story turns, and duck it under every line.
Do not ask a video model to invent a fight. Start from a reference fight that already works, strip it down to motion only (a faceless 3D motion-capture version, so no actor, costume or location carries over), and give the model your characters and setting separately. The model then follows the reference's choreography, timing, blocking and camera moves with your cast. Keep each reference between 4 and 15 seconds, match your setting to the reference's shape, cast the hero and the opponents separately, and check the result side by side with the reference before you cut it in.
AI video models generate every clip from scratch, so the second shot of a scene rarely matches the last frame of the first: the room, the extras, the props in each hand and the blocking are all invented again. Cutting away hides it and first frames narrow it, but the fix that holds across a whole episode is video continuation: give the model the end of the previous clip and have it continue from there.
Who is in the show: the roles a season needs, what each character card must carry, and what makes one survive fifty episodes.
Build the cast as a set of functions before you build people. A season needs a lead with one want and one wound, an antagonist ladder, an ally who will be compromised, and a rival — four speaking roles per episode at most. Each character card carries a want, a wound, a rule they will not break, one visible tell and a line nobody else would say; anything the camera must draw goes in the bible once.
Consistency is a production decision made once, not a correction applied later. Fix the face, the voice, the era and the wardrobe in the bible before the first render; give every character reference views and one voice; and write appearance notes that describe only what does not change. Everything you leave vague gets re-invented per shot, and re-invention is what the audience reads as a different person.
The first thirty seconds, the last five, and the turns that keep a viewer paying for the next episode.
A hook is a certain promise plus an uncertain process. In the first five seconds show a concrete image that promises the kind of satisfaction the series will deliver; in the next twenty-five, land three escalating blows on the lead's identity so the viewer is certain payback is coming and only curious how. Mystery without a promise is not a hook.
End each episode on a specific question, and answer that exact question in the first two beats of the next one. Then ask a new question. A cliffhanger that is not answered within ten seconds of the next episode is a broken promise, and viewers stop trusting the format after two of them.
Not one big twist but many small ones, nested like dolls: each reversal reveals another underneath and recasts the one before. Put the largest between 40 and 60 percent of the season, and make it change what the lead believes about the past, not what they want. The most-watched second of any episode is the one just before a reversal lands.
Working formulas for the genres that travel: revenge, werewolf romance, forbidden love, and the seven emotional frames.
Every short drama that travels runs on one of seven emotional frames, each a four-beat arc from a low to a specific release. Pick the frame before you plot, because the frame decides the order of events, the ending the audience will accept, and the one thing you must not add. Mixing frames mid-season is the most common reason a series loses its audience.
Revenge is the largest frame in Western vertical drama and the easiest to get wrong. The formula is four beats in strict order: swallowed humiliation, quiet accumulation of a new power, scores settled one by one from smallest to largest, and a public reckoning. No romance resolves before the last score is settled, and the person who hurt the lead is never forgiven on screen.
Werewolf romance in vertical format runs on a fixed cast web (the rejecting Alpha, the fated Alpha King, the underestimated Luna-to-be, the scheming false mate) and one of four plot formulas. Each formula is the same shape: rejection, exile, a stronger mate who recognises her, the ex's regret, and a public reversal where the false mate is exposed. The ex is never redeemed.
Billionaire romance in vertical format runs on a wealth gap made visible in the first shot, a fixed cast web (the cold CEO, the underestimated heroine, the rival who has the status she lacks, the family that looks down on her) and one of five formulas: contract marriage, mistaken identity, the hidden child, the substitute bride, and the secret heiress. Every formula pays off in public: the people who humiliated her watch him choose her, or watch her turn out to outrank them.
Mafia romance in vertical format runs on one promise: the most dangerous man in the city is dangerous to everyone except her. The cast is fixed (the don or heir, the woman pulled into his world, the family that wants her gone, the rival family, and a betrayer inside the house), and most seasons use one of five formulas: the alliance marriage, the debt, protective custody, the rival family's daughter, and the woman who came to bring him down. The violence stays mostly off screen. What viewers watch for is who he protects and what that costs him.
Hidden identity is the biggest male-lead genre in vertical drama. A man everyone looks down on (the live-in son-in-law, the security guard, the returning soldier, the delivery rider) is secretly the richest, strongest or most powerful person in the room, and the season is a ladder of moments where someone humiliates him and then finds out who he is. The cast is fixed (the hero, the wife or love interest who believed in him, her family, the rival who mocks him, and the people who know his real name), and the craft is in pacing the reveals so every episode pays one off without spending the big one too early.
Forbidden love works when the obstacle is a permanent, institutional rule rather than a temporary misunderstanding. There are five tracks that consistently perform: mentor and adult student, step-siblings, age-gap with a seasoned lead, workplace hierarchy, and rival houses. Each has one line that must not be crossed on mainstream platforms; write the tension from restraint, not from proximity to discovery.
What platforms buy, how many episodes they expect, and where AI-made series fit.
A vertical drama is a scripted series shot in portrait (9:16) for phones, told in episodes of one to two minutes, usually 50 to 80 of them, each ending on a cliffhanger. "Micro drama", "short drama", "mini drama" and "vertical series" all mean the same thing. People watch them in apps like ReelShort and DramaBox, on TikTok, YouTube and Instagram.
TikTok Drama Center (drama.tiktok.com) is TikTok's business portal for short drama supply. A verified company uploads a finished series of 30 to 300 episodes; TikTok reviews it, subtitles it into 17 languages, posts it region by region, and pays monthly from a share of ad revenue, a share of in-app purchases and platform incentives. Below: who can join, specs, review, regions and money, from the help center and our three live series.
Yes, with disclosure. Major short-drama apps already list AI-produced series in their catalogues at full season length, social platforms distribute labelled AI content, and China's 2026 rules require per-episode AI disclosure rather than banning it. What platforms reject is undisclosed synthetic media and content that fails their ordinary rules on likeness, age and safety. Label every episode, own every face and voice, and the format is open to you.
Dub the first ten episodes into one language, publish them on accounts that serve that market, and compare how far viewers watch against the original before you dub the rest. Keep one voice per character for the whole season, check who is speaking before anything is dubbed, lip-sync the dialogue, and burn in subtitles in the new language. Count cost and decisions by the episode, the same unit you sell.
Seedance API prices across providers, moving an integration between them, and which model tier fits which shot.
Seedance API pricing runs from about $0.12 to $0.24 per second of 720p video for Seedance 2.0 Fast and about $0.20 to $0.47 for Seedance 2.5, depending on where you buy it (public rates as of 2026-10-13). BytePlus ModelArk sells it direct in output tokens; resellers such as fal, Replicate and Dawnwood Models sell it per second. The table lists each provider's 720p rate and how it bills.
From BytePlus ModelArk, change two lines — the base URL to https://models.dawnwood.ai/v1 and the key to a Dawnwood token — because Dawnwood Models accepts the ModelArk request body, path and task object as they are. From fal, the request is reshaped once - the prompt and image URLs become items in a content array, aspect_ratio becomes ratio, duration becomes an integer - and polling moves from fal's status and response URLs to one GET on the task id. Both paths are shown side by side below.
Dated, short posts on model launches, price changes and market numbers. Each one is later folded into an evergreen guide.
Seedance 2.5 Flash is a faster, cheaper tier of Seedance 2.5 on BytePlus ModelArk (model id dreamina-seedance-2-5-flash-261115). It lists at US$7.0 per million tokens without video input, about 35% below Seedance 2.5, and outputs 480p or 720p clips of 4 to 30 seconds. In our one test it finished in 76 seconds against 139, with flatter detail.