How to Make an AI Comedy Series With Recurring Characters

You can make an AI comedy series with a recurring character by locking one face to a small set of reference images before you write a single episode, then generating every episode from that same reference set. On VdoBloom, the Short-Drama Series tab keeps a fixed cast and continues each episode from the last frame of the one before it, while HappyHorse 1.1 Reference-to-Video generates new scenes around your references so your lead looks like the same person in episode 1 and episode 6.

Updated October 2026.

Why the recurring character is the hard part

Comedy runs on recognition. A catchphrase is only funny if you recognise the person delivering it, and a running gag only escalates if the same deadpan face keeps walking into the same disaster. That is exactly what prompt-driven video models are worst at on their own: describe a woman in a mustard cardigan in one generation and a woman in a mustard cardigan in the next, and you get two different women. Faces drift, glasses change shape, hair length wanders.

Two ideas do most of the work. First, a locked cast — a small set of named characters and locations you reuse in every episode. Second, reference images: stills the model treats as a guide to who the subject is, rather than as one fixed frame to animate. Keep both fixed and a series holds together; let either drift and the audience notices.

The VdoBloom setup: a locked cast for a short-drama series

VdoBloom’s Short-Drama Series tab is built around a locked cast. You set a premise, a genre, a template and a style, choose three to ten episodes, and define one to four characters and one to three locations. Each character has a name, a fixed face-and-hair description and a fixed outfit; each location has a description and a lighting note. You can upload a photo for any of them, or let the tool draw a Character Sheet from your text description.

The workflow has a rhythm worth understanding before you spend anything. Creating the series bible is free. Writing the season is a free step, so you can rewrite the whole thing as often as you like without touching your balance. Making the missing reference images costs their own image-model price, and each episode is then billed as two real steps — a storyboard pass and a video pass — with the exact cost shown before each one and a failed step refunded.

Episodes are generated in order, and each one continues from the last frame of the previous episode. That is what gives a micro-drama its continuity: the cast, the sets and the closing frame all carry forward. Episodes are vertical (9:16), and a free binge cut stitches the finished episodes into one continuous watch once you have at least two.

Reference images are what keep a face consistent

The dedicated tool for cross-episode consistency is HappyHorse 1.1 Reference-to-Video. It works differently from ordinary image-to-video: instead of animating your upload as the opening frame, it treats one or more reference images as a guide to identity and style, then generates a new prompted scene around that subject. The same lead can be placed in a fresh scene every episode without any re-shoot.

The specs that matter for a series: clips run 4 to 15 seconds at any whole second, output is 720p or 1080p, and there are nine aspect ratios including 21:9 ultrawide and 9:21 vertical. Pricing is about 9 credits per second at 720p and 11.75 at 1080p, so a 4-second 720p clip is 36 credits. Content filtering is MODERATE, the standard tier, and reference images of real people should be ones you have the right to use.

The practical workflow is to gather two or three clean reference images of each recurring character — consistent lighting and angle help — and reuse that same set across every prompt in the series. For two-character comedy beats, the Seedance family also accepts two reference photos tagged @Image1 and @Image2 directly in the prompt, a lighter option when a scene only needs two faces held at once.

Six prompt examples you can paste in

The series below is a deadpan two-hander set in an office kitchen. Swap the character descriptions for your own, but keep them identical in every prompt — that repetition is half of what holds the look together.

1. Establish the recurring lead (HappyHorse 1.1 Reference-to-Video, 9:16, 8s, 1080p). “Using my three reference photos of the same character — early-30s woman, curly black hair in a loose bun, thin gold glasses, mustard oversized cardigan over a white tee — place her behind a cluttered office kitchen counter at 8am. She lifts a chipped mug, sips, then freezes mid-sip with a defeated deadpan stare straight to camera. Locked medium shot at eye level, soft cool morning light from a window camera-left, subtle handheld micro-drift, no zoom.”
2. Two-hander beat (Seedance 2.5, 9:16, 12s). “@Image1 (Priya, curly bun, gold glasses, mustard cardigan) stands at the far end of an office corridor. @Image2 (Dev, tall, buzzed hair, navy polo, lanyard) power-walks toward her holding a cake box like a trophy. She does not react until he is arm’s length away, then slowly slides one step sideways. Single continuous take, camera pans right to keep both in frame, flat office fluorescents, hold on Priya’s blank face for two beats after Dev leaves frame.”
3. The reaction insert (Seedance 2, 16:9, 5s). “Start from this still of Priya’s face and animate a tiny reaction: her eyebrows rise a millimetre, her eyes flick to camera-right, she blinks once, unimpressed. No head movement, no camera movement, shallow depth of field, soft office light. Keep her face, glasses and hair exactly as in the still — this is a recurring character, not a new person.”
4. The catchphrase (Kling 3.0 Omni, 9:16, 8s, native audio). “Priya, curly bun, gold glasses, mustard cardigan, stands at a counter holding a to-go cup. She looks at the camera and says flatly: ’The microwave is a personality test.’ She takes a sip and stares. Static medium close-up, cool office light, natural hand movement, eye direction to camera. Audio: dry deadpan delivery, quiet room tone, no music.”
5. The running gag (WAN 3.0, 16:9, 20s, native audio). “One continuous take across four desks: Priya walks into frame carrying a lunchbox, presses a single button on the office microwave, and walks straight back out. The camera dollies slowly left past three increasingly alarmed colleagues as a low hum builds. Warm afternoon light, steady dolly, no cuts. Audio: a rising hum and one long beep at the end.”
6. The season title shot (Veo 3.1 Fast, Reference Mode, 16:9). “Using these three reference images to hold the character and the set, generate a 16:9 establishing shot of the same office at sunset: Priya sits alone at her desk, back to camera, the room dark except for her monitor. Slow push-in toward the back of her head, no dialogue, cinematic and wide.”

Veo 3.1 Fast bills a flat 40 credits per generation and its Reference Mode takes three images at 16:9 only, so it is best kept for a single recurring title card rather than every episode. Kling 3.0 Omni generates audio on every clip with no way to switch it off, which is why the catchphrase prompt is written to be performed rather than narrated over.

Shot length, camera movement and staying consistent

Comedy timing is editing, and most models hand you short clips, so plan the cuts before you generate. Seedance 2 runs 4 to 15 seconds; Seedance 2.5 and WAN 3.0 both reach 30 seconds, which is enough for a full running-gag take without a single cut. An 8-second beat is a comfortable unit for a setup-and-punchline, because the deadpan pause has somewhere to live.

For camera movement, deadpan reads best locked down: a static medium shot with a micro-drift keeps the joke on the face. Use a slow dolly or orbit only when the gag escalates, and never throw a whip-pan into the middle of a recurring shot — the audience stores the geography of a set across episodes, and changing it breaks the recognition the format depends on.

Consistency is mostly discipline. Write the character description once, paste it verbatim into every prompt, and never let it drift. Reuse the same two or three reference images rather than a fresh photo each time. Keep the outfit line fixed even when the scene changes, since prompt-driven models regenerate costume detail from scratch on every run. And respect the limits: Kling’s prompts cap at 2,500 characters, so a tight recurring description beats a sprawling one.

Frequently asked questions

How many episodes can I make in one AI comedy series?

The Short-Drama Series tab lets you plan three to ten episodes per season, with one to four characters and one to three locations. Episodes are generated in order because each one continues from the last frame of the previous episode.

Do I need to upload photos of a real person?

No. You can describe each character in text and let the tool draw a Character Sheet, or upload your own photo for any character or location. Reference images of real people should be ones you have the right to use, and the tool asks for consent confirmation before it accepts uploads.

Which model keeps a character most consistent across episodes?

HappyHorse 1.1 Reference-to-Video is purpose-built for it, generating new scenes around a fixed reference set across nine aspect ratios. Inside the series tab, the Seedance family drives the episodes and holds the locked cast, with Seedance 2.5 supporting the longer 30-second beats.

How much does a comedy series cost to make?

The series bible and the season writing are free. References, storyboards and episode renders are each billed separately, with the cost shown before every step and a failed step refunded. A HappyHorse reference clip is about 9 credits per second at 720p. Full plan details are on the pricing page.

Can I reuse the same character in a different series later?

Yes — save the reference images and the exact character description, and start a new series with the same locked cast. The character consistency guide walks through building a reusable reference set.

Ready to try it? Open the Short-Drama Series tab, set your cast and premise, and the first season plan is free. If you would rather test the consistency idea on a single clip first, the HappyHorse reference walkthrough and the Seedance 2.5 guide both start from one reference image, and the reference-to-video pricing breakdown shows what a full season of references adds up to.