ByteDance opened a micro-drama factory, and here is how to use it

ByteDance built the thing every micro-drama producer has been half-assembling for two years: one window that writes the script, designs the characters, plans the shots, generates the video and cuts the result, then re-shoots the whole series for a different country. It is called the Short Drama Studio, it went live on October 26, 2026, and it is inside Jimeng, the company's Chinese app. The overseas brand, Dreamina, has no announced date for it.

That is the bad news, and it is the first thing to get straight, because every guide to this launch is written as if the studio were a download away. It is not. But the models behind it are another matter: Seedance 2.5 is available outside China through Dreamina and the BytePlus API, and the method the studio encodes is a method you can run today with three tabs and a folder. This is that method, with the costs and the access routes at each step.

What ByteDance actually shipped

Three house models and a post-production layer, according to the launch coverage and the pipeline description:

• DreamScript writes. It handles original scripts, adaptations, overseas versions and explainer cuts, reads a novel of up to 100,000 Chinese characters, and drafts up to 100 episodes in sequence while checking plot conflicts and character setup. It also runs what the Chinese coverage calls a hit diagnosis: it reads platform trends and hit logic to suggest structural fixes.
• Anycook is the image model, for character designs and scene styles.
• DreamShot is the storyboard model, with a canvas and a storyboard view, a 3D director's console for setting camera positions and movement before generation, and previews of key shots so you can see how they cut together.
• Post adds per-shot reshoots, voice and background separation with re-dubbing, automatic shot splitting, and an asset library where characters, costumes and locations are stored for reuse.

The feature with the most commercial logic is the remake and redraw tool: upload a finished drama, and the system maps its characters, settings, props and shot structure, then regenerates the visuals for a chosen region, style and aspect ratio while keeping the original's pacing and camera language. ByteDance inherited much of this from Xiaoyunque, the standalone short-drama platform it folded into Jimeng in October, as we reported at launch.

Getting access: three doors, only two open

Door one, the Studio: China only. Jimeng is at jimeng.jianying.com, and the Studio is inside it. A Chinese account is the practical requirement, the interface is Chinese, and the studio itself carries no published price. Jimeng's plans run from 393 yuan a year for the basic tier to 1,099 yuan a year for standard, with higher tiers, and the app is currently promoting Seedance 2.5 at 720p at 0.18 to 0.45 yuan a second, depending on tier, through October 28. If you are outside China and not set up for it, treat this door as closed for now.

Door two, Dreamina: open, minus the Studio. Dreamina is ByteDance's overseas brand at dreamina.capcut.com. It gives every account free credits that refresh daily, and it carries Seedance 2.5. Its live pricing page advertises a first-month rate from $0.01 a second on the basic plan, with paid plans settling at $0.05 to $0.08 a second; the $0.04 rate that circulated in September belonged to a campaign that ended on October 9. What it does not yet have is the Studio's one-window pipeline, so you assemble the equivalent yourself: a script model, an image model, a storyboard, Seedance, and an edit.

Door three, the API: open, and the one to use for volume. BytePlus ModelArk serves Seedance 2.5 with 4 to 30 second clips, 30 second single-take narratives, up to 50 multimodal references, and precise editing that swaps a product, face, style, lighting or background while keeping the motion. List pricing at 1080p was about $0.569 a second after the October 8 update, with resellers between roughly $0.59 and $1.16.

If you want the Studio's script and storyboard stages specifically, the honest substitutes are any strong reasoning model for DreamScript's job and any 3D tool, Blender included, for DreamShot's. Anycook has no announced overseas endpoint.

The format before the tools

None of this matters if the format is wrong. A micro-drama is not a short film; it is a serial engine:

• Episodes of 60 to 120 seconds, with the major platforms clustering near 90.
• 60 to 100 episodes a season, which is 90 to 150 minutes of finished runtime.
• Vertical, phone-first, subtitled, because most of it is watched muted in a feed.
• A turn at the end of every episode, not a resolution: the cliffhanger is the product.
• A small cast and one location per scene, which is a producer's rule of thumb rather than a platform number: it is what a model can hold consistent, and what a short shoot can afford.

Write to that shape and the tools behave. Write a 12-minute short and you will spend the budget fighting the format.

Step 1: premise, then the ten-episode test

Pick a premise with a built-in engine: revenge, hidden identity, a contract marriage, a comeback, a debt. The engine matters more than the idea, because you need 60 turns out of it.

Then write the first ten episodes before you generate a single frame. Ten episodes is about fifteen minutes of screen time, it is enough to hear whether the engine turns, and it costs you an afternoon instead of a generation budget. Each episode needs: the situation, the reversal, and the line that lands on the cut.

DreamScript is built to do this part at scale, and any capable model can do it at human scale. What no model can do is decide what the show is about, so decide that first.

Step 2: the character bible, and the locked reference

This is the stage that decides whether your series looks like one show or sixty auditions, and it is the stage the Studio's asset library exists to enforce.

For each character, lock one character sheet: front, three-quarter and profile, in the signature wardrobe, on a neutral background, at the aspect ratio you will shoot. Keep the file. It becomes the reference image for every shot that character appears in.

• Name the character in the filename, and keep a wardrobe variant per arc.
• Keep a location sheet the same way, one per set, with a fixed camera height and lens.
• Write a five-line style bible: genre, era, palette, lens language, and what the show never does. Dreamina's own film guidance calls this a rulebook, and it is right: it is the only thing that keeps separately generated shots in the same world.

Step 3: the shot list is the production

Convert each episode into five to nine shots of four to eight seconds. That is what a 90-second episode actually is, and it is the unit you generate, review and re-roll.

Write every shot as one line with five parts: subject, action, camera, light, duration. For example: "Mei, in the red coat, steps off the kerb into the rain; camera low and tracking right; sodium streetlight and neon reflection; five seconds."

Two rules from the people who do this at volume:

• Generate stills before motion. A still is cheap to fix and a clip is not.
• Use the first and last frame where a shot has to land on a specific composition, which is most of them in a serial.

Step 4: generate, then re-shoot the shot, not the episode

Seedance 2.5's 30-second single take is the tool for dialogue scenes: one generation holds the pacing and the eyelines. For everything else, generate 4 to 8 second clips and cut them.

The economics of re-rolling are the whole game. A bad clip is not a disaster if you can regenerate that shot; it is a disaster if a continuity error forces a re-render of the scene. The Studio's per-shot reshoot is exactly this, and you can run it manually: keep the shot list as a spreadsheet, one row per shot, with its prompt, its reference image, its takes and its chosen file.

Step 5: voice, sound and the dub

Generate dialogue as audio first and feed it in as a reference where the model accepts one, or generate silent and dub in post. The Studio's post layer separates voice from background and regenerates the dub in place, which is the workflow to copy: never bake a performance you might re-cast.

• Record or synthesise one voice per character and keep it consistent across the season.
• Keep music and effects on separate tracks, because a platform note will ask you to change one of them.
• Burn in subtitles. Most of your audience is watching with the sound off.

Step 6: the edit is where a series is won

Cut in 9:16, and cut to the turn. The first three seconds of episode one decide whether there is an episode two, so open on the inciting image, not on a title card. Keep the cuts on the beat, keep the subtitles above the platform's interface furniture, and end every episode on the line that makes the next tap inevitable.

Step 7: localize, which is the actual business

The remake tool's promise is that a finished season can be re-shot for another market without re-writing it: same rhythm, same camera language, new faces, locations and styling. You can approximate that manually by regenerating each shot with a region-specific location and wardrobe reference and the same shot list, which is why the shot list is the asset that matters. The script is portable. The bible is portable. The renders are not.

What it costs

These are the published rates, not a quote. A 90-second episode is 90 seconds of generated video, before retries.

• BytePlus ModelArk at 1080p, list: 90 × $0.569 = about $51 an episode. A 60-episode season, 90 minutes, is about $3,070.
• Dreamina's advertised first-month rate: 90 × $0.01 = about $0.90 an episode, or about $54 for the season. On the paid-plan range of $0.05 to $0.08 a second, the same episode is $4.50 to $7.20, or $270 to $432 for the season.
• Jimeng at 720p, promotional: 90 seconds at 0.18 to 0.45 yuan a second is 16.20 to 40.50 yuan an episode, or 972 to 2,430 yuan for the season.

Then multiply for reality. Assume two to three takes per shot on a first season, which multiplies the video line by two or three, and add voice, music and your own time. The promotional rates are promotional: both the Dreamina offer and Jimeng's per-second price are time-limited, and the API list price is the number to plan against.

The traps, in the order they will find you

• Identity drift across shots. Fix: the character sheet, referenced on every generation, and a continuity pass at thumbnail size before you watch anything at full size.
• Hands, text and small props. Write shots that do not depend on them, and cut around the ones that do.
• Lip-sync on close dialogue. Use the 30-second single take for those scenes, or shoot them as over-shoulder and dub.
• Rights. Real faces, real brands and real music are three separate clearances, and a generated likeness of a real person is not a loophole.
• Platform rules. Label AI-generated material where the platform asks, and read the rules of the app you are publishing into rather than the ones you published into last year.
• The Studio is not here yet. Do not plan a season around a tool with no overseas date and no published price.

The seven-day version

Day one, premise and ten episodes. Day two, the character and location sheets. Day three, the shot list for episodes one to five. Day four, stills. Day five, motion for the first two episodes. Day six, voice, sound and the cut of episode one. Day seven, watch it with someone who does not like you very much, and fix the first three seconds.

That is the whole factory. ByteDance has put a very good one in a box in China; the method in the box is free, and it works with the tools you can already open.