How to Make an AI Short Drama with Seedance 2.5, Wan 3.0 and Kling 4.0

A vertical short drama is a story told in short episodes, each ending on a hook that makes the viewer swipe to the next one. With AI video models, the hard part is not one nice clip but production: the same faces across twenty shots, dialogue that lands, action that reads on a phone, and an episode that survives the edit.

This guide walks through the full workflow we recommend on WaveSpeedAI, from a one-page script to an exported 9:16 episode. It uses three models that cover the job between them: Seedance 2.5, Wan 3.0 and Kling 4.0.

The five-step workflow

Every episode goes through the same five steps:

1. Script - a short, timed script with one location per scene and a cliffhanger at the end.
2. Character references - one clean reference image per character, reused in every shot.
3. Storyboard and shot list - each scene broken into shots of 4 to 10 seconds, each with framing, action and sound.
4. Generate each shot - pick the model that fits the shot, generate at draft resolution, then re-render the keepers.
5. Voice, edit and deliver - assemble, check the vertical safe zones, add a title card, export 9:16.

The rest of this post goes through each step with real prompts and settings. The example series is an original one: Night Bloom, about Mara, a night-shift florist, and Jun, a courier who keeps delivering flowers to an address that does not exist.

Step 1: Write a script that is built for shots

AI models generate shots, not scenes, so write the script the way you will generate it. Keep every scene in one location, give every line of dialogue to a named character, and keep each episode to roughly 60 to 90 seconds of screen time. A one-minute episode is usually 8 to 14 shots.

Use consistent scene headings, character names in capitals and dialogue under each name. A formatted script makes the next step mechanical: every scene heading becomes a location reference and every character cue becomes a character reference.

EP 01 - "WRONG ADDRESS"

INT. FLOWER SHOP - NIGHT
Rain on the window. MARA (late 20s, copper bob, green apron) trims stems.
The bell rings. JUN (30s, yellow rain jacket, delivery bag) steps in, soaked.

JUN: Same address as last night. Fourteen Orchard Lane.
MARA: There is no Fourteen. The street ends at twelve.
JUN: Then who keeps signing for them?

He turns the delivery slip around. The signature reads: MARA.

The episode ends on a reveal, not a summary; plan that last shot first. For the craft of it, see our guide to writing a vertical AI short drama script that ends on a cliffhanger.

Step 2: Lock your characters with reference images

Consistency is decided before you generate a single video. Make one front-facing, evenly lit reference image per character on a plain background, plus one image per location. Write each character’s name, wardrobe, hair and distinguishing details in one short block, and paste that exact wording into every prompt.

All three models accept references, in different ways:

• Seedance 2.5 Text-to-Video takes up to 30 reference_images, plus reference videos and reference audio. You bind them in the prompt as @Image 1, @Image 2 in upload order.
• Wan 3.0 Reference-to-Video takes up to 10 reference images, 5 reference videos and 5 reference audios, referred to as Image 1, Video 1, Audio 1.
• Kling 4.0 Reference-to-Video takes up to 10 images, 5 videos and 3 audios, referred to as @Image1, @Video1, @Audio1.

The deeper technique, including last-frame continuation, is in keep the same character across AI video shots.

Step 3: Turn the script into a shot list

In a spreadsheet or on paper, give each shot one row: shot number, duration, framing, action, dialogue and sound. One action per shot: “and then” usually means two shots.

Shot	Length	Framing	Action and sound
1	4s	Wide, rain outside the shop window	Neon sign flickers, rain ambience
2	5s	Medium on Mara	She trims stems, the door bell rings
3	6s	Medium two-shot	Jun steps in and says his line
4	5s	Close-up on Mara	She answers, puzzled
5	4s	Insert, the delivery slip	The signature reads MARA

Step 4: Pick the model per shot

You do not need one model for the whole episode. Use each where it is strongest, according to what its model page lists:

	Seedance 2.5	Wan 3.0	Kling 4.0
Good for	Dialogue shots with native audio, reference-heavy scenes	Low-cost drafts, inserts and long takes	Long single takes with action and camera movement, keyframed sequences
Inputs	Text, up to 30 reference images, reference video and audio; start and last frame in image-to-video	Text; first and optional last frame; image, video and audio references	Text; first frame, last frame and up to 8 in-between keyframes; image, video and audio references
Duration	4 to 30 s	2 to 30 s	3 to 30 s
Resolution	480p, 720p, 1080p, 4k	480p, 720p, 1080p	720p, 1080p
Audio	Native synchronized audio (on by default)	Optional audio	Synchronized stereo audio
List price, 720p	$0.36 per second	$0.10 per second	$0.157 per second

Prices are model-page list prices at the time of writing; the Generate button shows the current price. For a full side-by-side, read Seedance 2.5 vs Wan 3.0 vs Kling 4.0.

A practical split for Night Bloom:

• Shots 1 and 5 (establishing shot and insert) on Wan 3.0 at 720p. They have no dialogue and are cheap to redo.
• Shots 3 and 4 (dialogue) on Seedance 2.5 with both character references, audio on.
• Any chase or fight in later episodes on Kling 4.0, where a single take of up to 30 seconds leaves room for the whole move.

A dialogue shot on Seedance 2.5

On Seedance 2.5 Text-to-Video, upload Mara’s sheet as image 1, Jun’s as image 2 and the shop as image 3. The model page recommends a three-layer prompt: a one-sentence summary, a timed shot breakdown, then consistency notes.

Jun from @Image 2 steps into the flower shop from @Image 3 and speaks to Mara from @Image 1, cinematic night drama, medium two-shot, slow push-in.
0-2s: The door bell rings; Jun steps in, rain dripping from his yellow jacket.
2-6s: He holds up a delivery slip and says, "Same address as last night. Fourteen Orchard Lane."
Warm practical light from the shop, cool blue rain outside the window. Rain ambience and the door bell, no music, no subtitles.

Settings: aspect_ratio 9:16, resolution 480p for the draft and 720p for the keeper, duration 6, generate_audio true. For more on spoken lines, see AI lip sync dialogue scenes with Seedance 2.5.

An establishing shot on Wan 3.0

Wan 3.0 Text-to-Video is the cheapest of the three per second and accepts 2-second clips, which makes it good for inserts and establishing shots.

A small flower shop on a rainy city street at night, seen from across the road. A pink neon sign reading "BLOOM" flickers above the window. Rain streaks through the light, puddles reflect the neon. Static wide shot, slight drift to the right. Soft rain and distant traffic.

Settings: aspect_ratio 9:16, resolution 720p, duration 4, generate_audio true. For a fixed start and end, use Wan 3.0 Image-to-Video with a first frame and optional last_image.

An action beat on Kling 4.0

Kling 4.0 Text-to-Video generates any length from 3 to 30 seconds in one pass. Its prompting guide asks you to name the subject, action, setting and camera, describe the sound, and for longer clips, describe how the scene develops from beginning to end.

A courier in a yellow rain jacket sprints down a narrow wet alley at night, chasing a delivery van. He vaults a stack of crates, slips on the wet cobbles, catches himself on a wall and keeps running as the van turns the corner. Handheld tracking shot from behind, then the camera swings ahead of him for a low-angle shot as he passes. Rain, splashing footsteps, the van's engine fading away.

Settings: aspect_ratio 9:16, resolution 720p, duration 8. The multi-shot action scenes with Kling 4.0 guide shows how to chain several of these into one sequence.

Draft cheap, then re-render

Generate every shot at the lowest useful resolution first (480p on Seedance 2.5 and Wan 3.0, 720p on Kling 4.0), cut a rough episode, and only re-render the shots that survived the edit at 720p or 1080p. Per-second pricing means a rough cut costs a fraction of the final one. Keep the prompt identical between draft and final render; change only the settings.

Step 5: Voice, edit and deliver in 9:16

Once the shots are chosen:

• Continuity. If a shot has to start exactly where the previous one ended, export its last frame as a still from your editor and use it as the first frame of the next image-to-video generation.
• Dialogue fixes. If a line came out wrong but the picture is right, record or synthesize the corrected line and re-sync it, or rebuild the close-up with Seedance 2.5 Talking Avatar, which animates a portrait from an audio recording.
• Safe zones. App interfaces cover the top, bottom and right edge of a vertical video. Keep faces and captions in the center and preview on a phone before publishing. Details in vertical 9:16 AI video settings for Shorts and Reels.
• Title card. Design one 9:16 title card with the series title and episode number and reuse it so every episode opens the same way.

Export at 1080x1920 from 1080p shots or 720x1280 from 720p shots, and render final shots at the resolution you deliver.

Common mistakes

• Writing scenes, not shots. One prompt that tries to cover a whole scene produces cuts or dropped beats. One action per shot.
• Changing character wording. “Copper bob” in one prompt and “short red hair” in the next is a different person to the model. Copy the character block verbatim.
• Generating 16:9 and cropping. Cropping a widescreen shot to 9:16 loses most of the frame and usually cuts off a face. Set aspect_ratio to 9:16 at generation time, or give image-to-video a vertical first frame.

FAQ

Which model should I use for an AI short drama? Use more than one. Seedance 2.5 is a strong default for dialogue shots because it generates synchronized audio and takes many reference images. Wan 3.0 is the lowest-cost option for drafts and inserts. Kling 4.0 suits long single takes with action, and its image-to-video accepts in-between keyframes.

How long can each clip be? Up to 30 seconds on all three: Seedance 2.5 from 4 to 30 seconds, Wan 3.0 from 2 to 30, Kling 4.0 from 3 to 30. For drama, 4 to 10 seconds per shot cuts better than one long clip.

Can I generate vertical 9:16 video directly? Yes. All three text-to-video models list 9:16 as an aspect ratio. For Seedance 2.5 image-to-video, the output follows your start image, so give it a vertical image.

How much does a one-minute episode cost? It depends on the model mix and the number of retakes. At the list prices above, one minute of final 720p footage is $6.00 on Wan 3.0, $9.42 on Kling 4.0 and $21.60 on Seedance 2.5, before retakes. Drafting at 480p cuts the cost of exploration.

Ready to start? Open the video generator with Seedance 2.5, Wan 3.0 or Kling 4.0 preselected, or direct shots with camera presets in Cinema Studio.