November 21, 2026
AI tools for vertical drama: what to use for casting, dialogue and the final cut
description: "Compare Kling, Seedance, Runway, Vidu, Google Flow and local Wan models for a scripted vertical series. Includes a 30-second scene, a dubbing workflow and a cost-per-finished-minute breakdown." slug: "ai-tools-for-vertical-drama" lang: "en" date: "2026-11-21"
AI tools for vertical drama: what happens when you have to change one line?
John Chua spent about 45 hours making a two-minute animated short. In one take he asked a character to wave. She produced a sword. Chua cut away just before the blade appeared. A different shot came back in the wrong visual style, so he used it as a kind of flashback. Both shots earned their place in the edit. Had the sword covered a clue the audience needed to see, he would have had to go again. Chua’s interview with LSU
For a scripted vertical series, the first question to ask of an AI video tool is which part of the scene can you afford to lose? If the scene turns on an actor swallowing a word, start with Runway Act-Two and a recorded performance. If the audience has to recognize the same key three episodes later, try Vidu with a prop reference. If someone must stop at the door on a particular beat, try Kling’s timed keyframes. A phone rehearsal worth keeping gives Seedance or Luma a different starting point. Act-Two dialogue workflow · Vidu references · Kling keyframes · Seedance · Luma Modify
Which AI filmmaking tool fits your series?
| What you bring to the project | Tool to try first | What to check in the cut |
|---|---|---|
| A script that has yet to be staged | LTX Studio breaks a script into scenes and shots and extracts characters, locations and objects. Toonflow offers a self-hosted workspace for a recurring cast, shot boards and a choice of generation models. LTX storyboard workflow · Toonflow project | Cut the boards against a scratch track. With captions off, can you tell who spots the clue first? |
| A recurring lead, costume or distinctive prop | Runway Gen-4 Image References for character stills and opening frames. Vidu Reference to Video to generate a shot from character, prop and location images. Runway References · Vidu reference workflow | Put a front view, a profile and a backlit medium shot side by side. Check the coat and the object in the actor’s hand as well as the face. |
| A move that has to hit a mark | Kling 4.0 lets you put keyframes at specific times, including mid-shot. Its Keyframes mode takes no additional reference assets, so put the approved cast and set into those frames first. Kling 4.0 guide | Watch the approach, the stop and the turn on a phone. Does the actor stop on the correct side of the threshold? |
| A good rehearsal filmed on a phone | ByteDance Seedance 2.5 takes video references for framing and movement, generates picture and sound together and supports up to 30 seconds per generation. Luma Ray3 Modify can change a character’s appearance in source footage. Seedance model page · Ray3 Modify guide | Pick the moment a face passes behind a shoulder or an actor shifts their weight. Does the new shot keep it? |
| A portrait-format location shot with sound | In Google Flow, Veo 3.1 Ingredients to Video supports image references and native 9:16 output. Veo’s portrait mode · Audio and editing in Flow | Listen for the line, the room and the latch separately. Keep a clean line recording if a rewrite is likely. |
| A two-hander carried by interruptions and pauses | Runway Act-Two transfers a recorded performance to a character. Runway’s two-person tutorial makes a separate performance pass for each actor and composites the results in an editor. Runway’s two-character example | Stay on the listener while the speaker finishes. Which word makes the listener look up? |
| A local GPU and a recurring type of shot | Wan2.2 TI2V-5B has downloadable weights and inference code. Its documented single-GPU 720p-class example uses offloading and needs at least 24 GB of VRAM. That figure belongs to this particular 5B setup. Wan configuration and requirements | Time a close-up at delivery settings. Count how many seconds survive the edit before committing hardware to a season. |
For a solo motion-comic series, a character sheet, Vidu and an editor you already know will take you through a pilot. A production workspace starts to pay off when someone asks which coat the lead wore in episode three. You can host Toonflow yourself and still pay for calls to an external video model. For a live-action-style family drama, put time into performances and reaction shots. For fantasy or suspense, test the one image you could never afford to shoot and see how much screen time it earns.
Runway put 11 models through the same ten briefs and scored Seedance 2.5 highest. Those briefs included a lipstick ad, a property tour and a dance. Borrow the head-to-head setup for a mystery pilot, then swap the lipstick for your recurring prop. Can you cut from the key in the sister’s hand to her brother’s view of it and recognize the same knot of red thread? Runway’s published test and briefs
A 30-second vertical drama, shot by shot
An original test scene. A brother and sister come home after their father’s funeral. She finds a brass key in her pocket. Her brother recognizes it. He put that key in their father’s coffin. Before either of them opens the door, a familiar voice calls from inside: “Wipe your feet.”
| Time | What the audience sees | How to make it |
|---|---|---|
| 0–4 s | Muddy shoes on a doormat. The sister reaches into her coat. | Animate a portrait-format opening frame. Give the mat enough space on screen. It sets up the last line. |
| 4–9 s | A brass key in her palm, red thread knotted through its bow. | Give the key an insert of its own. Photograph or draw one approved version and use it for every subsequent shot. |
| 9–15 s | Her brother looks down. “I buried that key with Dad.” | Record the line first. Have him look up only on “Dad.” That change of eyeline is worth a separate performance take. |
| 15–21 s | Her fingers close. Neither touches the door. The latch turns on the other side. | Cover the hand and the door separately. Lay in the latch sound. Watch on mute to make sure both siblings are visibly clear of the lock. |
| 21–30 s | Hold on the siblings. Their father speaks through the door. She looks down at her shoes. | An offscreen voice is all the father needs. Save the performance pass for her downward glance, and leave a beat at the end. |
Record three sounds before generating those shots: the brother’s line, the latch, the father through the door. Lay them under five stills. You can hear immediately if Dad comes in too early. The audience needs a moment to register that neither sibling touched the lock. LTX on cutting an animatic
Continuity includes where people stand. Alongside the character portraits, keep a two-shot with the sister on screen left and her brother on screen right. She looks right to meet his eye. He looks down and left at the key. Perfect likenesses with mismatched eyelines can make it look as though a third person has joined them on the landing. Runway’s sketch references let you indicate gaze and object placement. The ordinary rules of screen direction still do the work between shots. Sketch references · The 180-degree rule · Eyeline matches
For the insert, a useful image-to-video prompt is short: Locked camera. The palm opens slowly. The key lies flat, red thread hanging over the thumb. The hand comes to rest. The starting frame already supplies the metal, the colour and the framing. If opening the hand keeps producing a second key, start with the palm open. Carry the rustle of the coat over the cut. The viewer needs to recognize the key. The pocket business can go. Runway’s image-to-video prompting guide
How to lip-sync two characters and dub the same scene
Cut to the sister while her brother is still speaking. It gives her time to react and buys a little room in a second language. The Chinese line “这把我放进爸棺材了” becomes “I buried that key with Dad.” In a close-up of the brother, his lips have to carry every syllable. On the sister, the word “Dad” can land on her fingers closing around the key. That cut works in either language.
Three performances, three different revision bills. Dad is behind a door. A new recording or a text-to-speech pickup changes only his audio. The brother’s mouth is visible. Lock his recording first, drive his face from that performance and redo the visibly mismatched portion for the dub. The sister’s most important line is silent. Keep the take with the right glance. If a read has the right timing but needs a different vocal quality, ElevenLabs Voice Changer accepts an existing recording. Use an authorized voice and audition the result at the same edit point as the original. Act-Two performance capture · ElevenLabs Voice Changer · Text to Dialogue
Deliver a clean picture master and separate dialogue, music and effects. Adobe Premiere can translate an existing caption track. CapCut International also supports bilingual captions. Have someone who knows the story check names, family relationships and deliberate ambiguities. Then play the last shot with captions on. A subtitle that appears half a beat early can give away who is behind the door. Premiere caption translation · CapCut bilingual captions
How much does a finished minute of AI drama cost?
A video API bills for the footage it generates. An editor may keep only a few seconds of each take. On November 21, 2026, Runway’s developer rate card listed credits at $0.01 each, Gen-4 Turbo at five credits per second, Gen-4.5 at 12 and Veo 3.1 with audio at 40. These are prices on one API, in US dollars. The models differ in capability, including sound. Runway developer pricing
| 60 seconds of generated footage in the final edit | Per generated second | Shoot 180 s, keep 60 s | Shoot 360 s, keep 60 s |
|---|---|---|---|
| Gen-4 Turbo | $0.05 | $9.00 | $18.00 |
| Gen-4.5 | $0.12 | $21.60 | $43.20 |
| Veo 3.1 with audio | $0.40 | $72.00 | $144.00 |
Generation cost per finished second = video generation spend divided by the seconds of generated footage used in the final cut. The 3:1 and 6:1 ratios above are budgeting scenarios. Add character art, voice sessions, music, captions and editing labour as separate lines. If you shoot 20 seconds of an episode on your phone, charge them to live action and leave them out of the denominator.
For a pilot, give each candidate the same opening frame and line. Buy three takes apiece. Drop every take into the sequence and log how many seconds work and why you cut the rest. A face changed. A hand grabbed the wrong object. A reaction came one word late. At $0.05 per generated second, keeping one second in six costs $0.30 per usable second. At $0.12, keeping one in two costs $0.24. A preference leaderboard can narrow the field. Your prop and your dialogue settle the budget. Artificial Analysis blind-testing methodology · MovieBench on cross-scene consistency
When is it easier to shoot the real thing?
Producer Chen Kun described a ten-person team on the Chinese fantasy series 山海奇镜之劈波斩浪, with still-image generation and image-to-video work assigned separately. The appeal of fantasy was practical: the spectacle was expensive to achieve on a conventional microdrama budget. Eliza McNitt’s ANCESTRA made a different use of the technology. The short mixes live action with Veo footage. For a view from inside a hatching crocodile egg, the team generated elements with Veo and Imagen, then assembled the shot with traditional visual effects. Chen Kun interview, in Chinese · Yicai’s reporting · Inside the making of ANCESTRA
Keep the voice and likeness releases, music permissions, source assets and final edit together for delivery. YouTube asks creators to disclose realistic AI-generated or meaningfully altered material during upload. For releases on Chinese online content platforms, Article 10 of the national labelling measures requires users publishing generated or synthetic content to declare it and use the service’s labelling function. YouTube disclosure guidance · Chinese labelling measures
Back on the landing, an actual brass key and five minutes with a phone might finish the insert. Spend the saved time recording Dad twice. In one take he sounds stern. In the other he barely thinks about the words. He has said them every day of his daughter’s life. Keep whichever one makes her look like a kid coming home.

