If you have watched a short drama and thought you could make one, this is the most realistic moment that thought has ever had.

Vigloo released a 22 episode series called Bloodbound Luna in March 2026 with video, graphics, and voice acting all produced by AI. According to The Guru, fewer than ten people finished it in eight weeks. A live action short drama averages around 150 million won and takes two to three months, so the figures PD Journal reported for AI production, roughly one tenth the cost and one to two weeks, are not an exaggeration.

The hard part is how. Feeding one prompt into a video model does not produce a drama. The same character has to appear identically across cuts and episodes, dialogue has to be audible with matching mouth movement, and each episode has to pull the viewer into the next one.

This guide covers four stages in order: planning, character sheets, video generation, and episode design. Video generation is written around R2V on Seedance 2.5, where you supply references and get video back. That changes the order of work compared with older methods.

Kakao Ventures estimates the Korean short drama market at roughly 650 billion won in 2026. Issue Insight reported a forecast of 1 trillion won by 2028 and around 1.5 trillion won by 2030.

Korea converts well. Sensor Tower places Korea fourth in the world by share of monthly short form drama app revenue, behind the United States, Japan, and the United Kingdom, with a record month in January 2026. Herald Business reports that pricing is per episode at roughly 300 to 700 won rather than a subscription, so episode count is revenue.

Platforms are already filling up with AI. Herald Business reports that roughly 40 to 50 percent of the short dramas Vigloo launched in the past month were AI works, and that AI accounts for 15 percent of its catalogue overall. Inside the popularity rankings the AI share is 30 percent in Korea and 80 percent in the United States. Viewers are already choosing AI titles.

Broadcasters have crossed over too. A Korea Creative Content Agency survey put generative AI adoption in the broadcast and video sector at 3.2 percent in the fourth quarter of 2024 and 31.9 percent a year later, a tenfold rise. SBS produced a three minute action sequence for the drama Kim Bujang entirely with AI.

The cost structure is completely different. A Cine21 feature puts Korean live action short drama budgets between 30 million and 350 million won, averaging around 150 million. A working writer quoted in it says you have to shoot within twelve hours and wrap in six or seven days to hold the budget at 100 million. PD Journal reports that AI production cuts cost to roughly one tenth and shortens the schedule from two or three months to one or two weeks.

Three revenue routes have opened. Vigloo opened a producer dashboard in March 2026 that exposes settlement details and per episode drop off. TikTok introduced up to six times the payout for Korean language content from April 2026, and ZDNet Korea reports that YouTube now permits monetisation of AI produced work.

Competitions have grown as well. The 2026 Korea AI Content Festival put up 160 million won in total prizes including 20 million won for the short form drama grand prize, open to individuals and teams alike.

Planning works the same whether you generate or shoot. Short drama runs on such a compressed rhythm, though, that standard drama craft does not transfer unchanged.

Fix the logline as one sentence first. StudioBinder defines a logline as a one or two sentence summary carrying both the premise and an emotional hook. It holds that four elements have to be present: protagonist, goal, inciting incident, and conflict.

Screenwriter John August published a reader submitted pattern on his blog: when the inciting incident happens, the hero struggles against an antagonistic force to reach a goal before the stakes land. By his own count only about 6 percent of loglines take that shape, so he does not call it a formula. His conclusion is that a logline is not plot but an arrow pointing toward the story.

Design each episode in four beats. The vertical drama writing guide from Filmustage splits an episode into hook, friction, spike, and button. Hook runs from 0 to 15 seconds, friction from 15 to 60, spike from 60 to 90, and button covers the last 5 to 10 seconds. Friction has to be conflict that actually appears on screen, described as physical or verbal conflict in the frame rather than emotional subtext or implied tension.

Scripts run one to two pages per episode. Filmustage puts a vertical drama episode at one to two script pages across a series of 60 to 100 episodes. Final Draft, the industry standard writing software, ships a dedicated micro drama template and a separate ReelShort vertical drama template. Both assume one to three minutes per episode ending on a cliffhanger, with episode elements built into the formatting.

Test the first three seconds with a freeze frame. The Real Reel vertical drama guide argues that if you stop the frame at three seconds, a stranger should understand the situation without backstory. Chinese short drama crews call this segment the explosion point. The official Hongguo short drama writing course defines the job of the first 30 seconds as stopping a scrolling thumb.

Decide cliffhanger placement during planning. EBSCO Research Starters defines a cliffhanger as a device that sustains suspense by leaving a story unresolved at a critical moment, and sorts the content into four types: shocking twist, terrible accident, running out of time, and impossible choice.

Where you cut matters more. Writing instructor September C. Fawkes treats the backbone of a cliffhanger as a turning point in story structure and sorts cut position into four options: just before the turning point, mid way through when the turn has several layers, just after it at the moment of now what, and immediately after a new conflict lands.

Real Reel adds that you should cut two seconds earlier than feels safe, on the grounds that most writers over explain their cliffhangers.

Map the full run by episode in advance. Tencent News describes Chinese microdrama practice as splitting hooks into large and small. Small hooks close every episode while large hooks sit at payment nodes such as episodes 30 and 60. On a 100 episode run, at least 95 episodes carry a small hook.

The first thing to break in a short drama is the face. If a different person appears every time the cut changes, the story does not hold. So the character sheet comes before any video: a reference document holding front, side, and back views plus a face close up on one image.

Generate the base image first. Create one character as a front facing full body shot. The Carat character sheet guide advises framing the shot as full body, since a waist up image leaves too little information for the sheet. On Carat, the Hyperreal Portrait app handles this step.

If the prompt stalls, one approach is to split it into five parts: subject, background, lighting, camera, and style. The order and worked examples are in the portrait guide.

Unfold the base image into a sheet. Upload the image to the character sheet generator and you get a turnaround sheet with front, side, and back views plus a face close up.

Four sheet types are available: default turnaround, expression sheet, pose sheet, and costume sheet. For drama, keeping a turnaround and an expression sheet per character is worth the effort. The step by step procedure is in building a character sheet.

Write the script side character notes too. Separately from the image sheet, record what the character wants, what blocks it, how they change, and what secret or wound they carry. The series bible items Final Draft lists are personality, background, strengths, flaws, relationships with other characters, and motivations. It stresses recording how the characters cause conflict for each other. Filler such as personal taste can be skipped when it never enters the story.

Check how many references the model accepts. A sheet helps only if the model can take more than one reference, since face and wardrobe usually need separate slots. Google Gemini 3 Pro Image supports up to 14 mixed references per the official documentation, with six object slots and five character slots.

This is the biggest change from older workflows. Feeding an image in as a first frame and extending it into video meant preparing a separate first frame for every cut. R2V takes multiple references at once and assigns a role to each, which changes the order of work itself.

What is R2V. BytePlus ModelArk documentation lists three task types for Seedance 2.5: text only, first frame or first and last frame, and reference based generation. A request becomes R2V as soon as it carries at least one reference asset. If the prompt carries editing intent the model routes it to video editing, and extension intent routes it to video extension.

You can pass up to 50 references. That is 30 images, 10 videos, and 10 audio clips. Video and audio are each capped at 30 seconds in total. The previous 2.0 generation allowed 15, so the ceiling more than tripled.

Audio can also be supplied on its own starting with 2.5. Usage based on 2.0 is covered separately in the Seedance 2.0 guide.

Assign roles to references in the prompt. References are numbered in upload order as Image 1, Video 1, Audio 1, and the prompt states in words what each asset governs. Examples from the official prompt guide read like the knight in Image 1, Images 1-2 are Character 1 and correspond to Audio 1, and Image 1 depicts the protagonist and uses the voice from Audio 1.

Upload the sheets from stage 2 directly. Put the turnaround sheet in as the first image and the expression sheet as the second, then bind them to one character in the prompt with a phrase like Images 1-2 are Character 1. For a second character, repeat the pattern as Character 2. Write the dialogue in curly brackets and one cut comes out. That is a full loop.

Do not write names into the images. Labelling a photo with a character name and then calling that name in the prompt is explicitly discouraged in the documentation. Characters get confused or duplicated. Numbers are more precise.

Subject counts have recommended limits. Per the official guide, image references hold up steadily with one to eight subjects, while nine to twelve is possible with reduced stability. Audio and video references are recommended at one to five subjects. Audio or video carrying a subject performs best at five to ten seconds.

The first route feeds model output back in. A Seedance generated video containing a face, or the last frame image from that video, can be supplied again as a reference. The conditions are that it came from the same account, was generated within 30 days, and has not been edited.

Output from another platform or another account does not qualify, and secondary editing or compression voids it. The second route designates a preset digital character by asset identifier, and the third registers a real person asset cleared through identity verification and consent.

Dialogue and audio arrive with the video. Audio generation is on by default, so voice, sound effects, and background music are produced in the same pass as the picture. Mouth movement is matched here too, which removes the separate lip sync step. The official tutorial assigns a distinct marker per sound type: parentheses for music, angle brackets for sound effects, curly brackets for dialogue, and black lenticular brackets for subtitles. Eleven languages are supported, Korean among them.

Set output specs like this. Length runs from 4 to 30 seconds, resolution covers 480p, 720p, and 1080p, and vertical 9:16 comes out at 1080 by 1920. Pure R2V does not lock aspect ratio or length to the input assets, so you set them. First frame, editing, and extension tasks do inherit the aspect ratio of their input.

Getting one piece out is where the real work starts. Per the Filmustage guide, a vertical short drama runs 60 to 120 seconds per episode across 60 to 100 episodes. It is a serial, not a single video.

One episode is several clips joined together. A single pass tops out at 30 seconds, so a 60 to 120 second episode takes two or three clips. Seedance 2.5 handles extension in the same interface, so you supply the previous clip as a reference and describe the scene that follows. A parameter for setting the task type directly reduces cases where the model misreads intent.

Lock your character assets early. Because of the 30 day limit, base assets expire as a run gets longer. Keep strong character clips stored unedited, and refresh them by generating a new video within each 30 day window. Source URLs stay valid for 24 hours, so download and store them.

Win the first ten episodes. The official Hongguo course treats episodes 1 through 10 as the golden stretch where the core attraction has to be established immediately. Episodes 11 through 30 carry the main development, and 31 onward pay off setups and resolve emotion. Twists are governed by the principle of being unexpected yet consistent with reason.

Episode internal rhythm is defined too. The industry summary relayed by Tencent News is conflict or reversal every 15 seconds, plot advancement every 30, and suspense in the final 10. The Hongguo course divides an episode into a first 30 seconds, a middle minute, and a final 30 seconds, holding that in the middle stretch every line and action should serve the plot.

Alternate emotions to control pacing. The same course frames the creator as a manager of viewer emotion, alternating scenes of opposing feeling such as anger, warmth, and curiosity. A single sustained emotion wears the viewer out.

None of the four stages is tied to a specific tool. Each stage does need a different model, though, which means moving between services. Carat handles that in one place.

Characters through to video stay in one conversation. Build the base character and the sheet, then pass the result straight to a video model as a reference. Model and quality tier are selectable, and Korean prompts go in as written.

The available video models are listed in the model catalogue. Beyond the Seedance family, several video models can be tried side by side in the same place.

Use the gallery when character prompts stall. The character sheet prompt page collects prompts for turnarounds, pose sheets, and expression sheets. Take one as is and change only the character details.

There is a place to watch finished work too. The series tab in the Carat app carries vertical short dramas, built so you can continue the story yourself after watching an episode. It is useful for seeing what state a 30 second prologue has to end in for the next episode to feel necessary.

With planning and character sheets ready, generation itself is quick. For a first attempt, most of the time goes into producing the base character and the sheets. For reference, the fully AI produced 22 episode Vigloo series took fewer than ten people eight weeks. An individual attempting a short series should expect several days for the first episode.

Build the base image first, unfold it into a character sheet, then supply the same sheet as a reference on every cut. The official guide discourages writing a character name into the image and calling that name in the prompt. Characters get confused or duplicated, so use numbers.

Seedance 2.5 produces video and audio in the same pass. Write dialogue in curly brackets in the prompt and voice and mouth movement are generated together. Korean is officially supported, so Korean dialogue can go in as written.

Not as a direct upload. Only three routes are open: feeding model output back within 30 days, using a platform preset character, or registering a verified real person asset.

Korean short drama platforms include Vigloo, Shortcha, K-Short, and TopReels. Short form channels such as TikTok and YouTube Shorts have monetisation open, and AI content competitions are increasing. Starting short to test response before extending the episode count is the safer order.

AI short form drama production comes down to four stages. Plan with a logline and four beats, lock the character with a sheet, generate video and voice together from references, then continue into episodes. Characters are what to reach for first. Lock one person properly and every later stage gets easier.