5-Step Workflow: How to Make Viral AI Short Dramas (60M Views)

2026-10-22 · case-study tutorial · a guest · recording, 19:59 · watch on YouTube

Topic: Video generation & shot direction (also: Pre-production & story, Assets, references & consistency, Prompting & AI assistants, Studio work & business) · mode: Tutorial · tags: prompts shared, fiction, time-bound offer

A presenter breaks down Higgsfield's accountant-and-fashion-director AI short drama (over 60 million views) built in one ChatGPT chat with the Higgsfield plugin: a familiar storyline and a hook-context-escalation-cliffhanger structure, Soul 2.0 faces with character sheets combined in GPT Image and edited in Seedream 5.0, a three-quarter location view, then six Seedance 2.5 scenes with the failed takes and their fixes, and three ways to monetise short dramas.

What is said

The parts of the video and what is said in each (★ — how much it matters to the video).

00:00:00 — The AI drama that does not look like AI

★★ a guest: The presenter says Higgsfield's AI short drama hit 20 million views in 24 hours, and that people recreating it get millions of views. — 00:00:00 · also in: This AI short drama got 60M views and doesn't look like AI
★★ a guest: The presenter says AI short dramas bring creators collaborations, followers and clients who pay for AI videos. — 00:00:00 · also in: This AI short drama got 60M views and doesn't look like AI

00:00:20 — The short drama played in full

00:02:21 — Step 1: pick a story people get in seconds; length and the four parts of a short drama

★★ a guest: The presenter puts the short drama's more than 60 million views mainly down to its ultra-realistic look, like a clip from a real film, which is unusual for AI short dramas. — 00:02:21 · also in: This AI short drama got 60M views and doesn't look like AI
★★ a guest: The presenter connects Higgsfield to ChatGPT and builds the whole short-drama episode in one ChatGPT conversation. — 00:02:21 · Higgsfield ChatGPT Plugin, ChatGPT
★★★ a guest: For a short drama, the presenter advises picking a situation people grasp quickly and offers three starting points: a Cinderella story, love against the odds and a rivalry. — 00:02:48
★★ a guest: The presenter advises aiming for one to three minutes for a short drama. — 00:02:48
★★★ a guest: The presenter builds a short drama's story around four parts that every short drama is built on: the hook, the context, the escalation and the cliffhanger. — 00:02:48

00:03:36 — Step 2: characters that keep their face, character sheets for Lucy and the cast

★★ a guest: The presenter says a successful short drama also needs memorable characters and a setting. — 00:03:36
★★★ a guest: The presenter uses Soul 2.0, called the best image model for ultra-realistic faces and locations, to avoid the typical AI face. — 00:03:36 · Soul 2.0
★★ a guest: The presenter describes Lucy to ChatGPT, has it generate options with Soul 2.0, and adds early-2000s lighting, colours and film texture to the description so she fits the story's world. — 00:03:36 · ChatGPT, Soul 2.0
★★ a guest: When choosing among generated character faces, the presenter advises skipping the ones where the face looks washed out. — 00:03:36
★★★ a guest: The presenter gives each reference a clear name ("Lucy Face") and uses the same names throughout the ChatGPT chat so it is clear which references belong in each scene and no older version is used by mistake. — 00:04:21 · ChatGPT
★★ a guest: For an outfit reference ("Lucy Outfit") the presenter describes the clothes in the chat, or uploads a reference image and says which details to use. — 00:04:21
★★★ a guest: The presenter's rules for a character sheet: a plain grey background, three views (full body front, back, face close-up) and the face erased from the full-body front view, so the video model has one clear face reference instead of using different faces in different shots. — 00:04:52 · character sheet
★★ a guest: The presenter needs two versions of Lucy, with a bun and glasses and, after the fall, with her hair down and no glasses, and so makes the first character sheet and edits it into the second. — 00:05:13 · character sheet
★★★ a guest: The presenter combines the face and outfit references into the character sheet with GPT Image, saying it works really well with references and preserves the details. — 00:05:13 · GPT Image, character sheet
★★★ a guest: For a character's second look the presenter has ChatGPT edit the character sheet with Seedream 5.0: hair down, glasses removed, face and outfit kept exactly the same. — 00:05:46 · Seedream 5.0, ChatGPT, character sheet
★ a guest: The presenter uses the same workflow for the rest of the cast: Raphael the director, Rebecca his assistant and Luca the stylist. — 00:05:46
★★★ a guest: For small details such as jewellery, the presenter generates a separate close-up with Soul 2.0 and attaches it with the face and outfit references when making the character sheet, as he did for Raphael's necklace. — 00:06:10 · Soul 2.0, character sheet
★★ a guest: For the casting models the presenter had Soul 2.0 generate the whole group in one image, with different faces and poses, all in black, and uses it as the reference for the entire casting scene. — 00:06:10 · Soul 2.0

00:06:37 — Step 3: build the whole set in one image

★★★ a guest: The presenter generates the casting-room location in a three-quarter view showing the table, the runway and the entrance, because with a straight-on view the video model must guess the rest of the room when angles change; showing more of the layout keeps the location consistent. — 00:06:39
★★★ a guest: The presenter gives ChatGPT the story outline and asks for a scene-by-scene script around the chosen cast and location, with dialogue, action and a generation prompt for each scene. — 00:06:39 · ChatGPT

00:07:10 — Step 4: script it, with the two skills

★★ a guest: The presenter attaches two skills linked in the description; the script skill has a hundred story ideas built in and picks the strongest one. — 00:07:22
★★ a guest: The presenter says the second skill writes the prompts (camera, lighting, cinematic work), is built for Seedance and raises the generation win rate. — 00:07:22 · Seedance

00:07:51 — Scene 1: start with the loudest moment; fixing the fall and Raphael's accent

★★★ a guest: The presenter opens a short drama's hook with the peak of the scene: the viewer sees the loudest moment without knowing how it came about and keeps watching to find out. — 00:07:52
★★★ a guest: In the context part of a short drama, the presenter keeps emotional swings coming every 20 to 30 seconds: a starting point, then hope, another hit, then a chance. — 00:07:52
★★ a guest: The presenter asks ChatGPT to generate each finished scene through Higgsfield with Seedance 2.5. — 00:07:52 · ChatGPT, Seedance 2.5
★★ a guest: In the first take of scene 1 the model's fall looked unrealistic and a little silly and the empty runway did not feel like a busy casting, so the presenter asked ChatGPT to rework the shot and add models waiting in the background. — 00:08:26 · ChatGPT
★★ a guest: In the presenter's Seedance 2.5 take, Raphael spoke with an American accent although a French accent was asked for. — 00:08:58 · Seedance 2.5
★ a guest: In the same take Rebecca's headshot ended up among the model photos, but the presenter kept the take's dynamic acting and camera work. — 00:08:58
★★ a guest: To fix the fall and the misplaced headshot, the presenter tells ChatGPT exactly how those moments should play out. — 00:08:58 · ChatGPT
★★★ a guest: To get an accent right in generated speech, the presenter describes how it sounds in detail; for Raphael's French accent he asked for a soft uvular R made at the back of the throat. — 00:08:58
★★★ a guest: The presenter gives each character a voice description, has ChatGPT carry it into every scene where that character speaks, and listens to the takes together to check the voices match. — 00:10:11 · ChatGPT
★ a guest: In scene 2 Lucy walks in with her paperwork, trips and catches Raphael's attention; the presenter generates it from the script and the chosen references. — 00:10:11

00:10:33 — Scene 2: the fall that changes everything

★★★ a guest: For a short drama's unexpected turn, the presenter starts from a feeling the audience recognises (Lucy fears she will get in trouble) and turns it into an opportunity, which makes the situation relatable. — 00:11:09
★★★ a guest: In the escalation part of a short drama, the presenter calls for an unexpected turn: new characters show up, the conflict grows, things happen the viewer does not see coming. — 00:11:35
★★★ a guest: In a Seedance 2.5 scene prompt the presenter describes both performances (Lucy apologises nervously, Raphael's boredom turns into curiosity) and spells out the movement: he gets up from the red chair, leaves the table and crosses the floor to her. — 00:11:35 · Seedance 2.5

00:11:37 — Scene 3: the escalation and describing both performances

★ a guest: In scene 4 Raphael makes the offer, and the presenter has ChatGPT generate a separate close-up of Rebecca calling Luca. — 00:12:35 · ChatGPT

00:12:37 — Scene 4: the offer and Rebecca's side-eye

00:13:16 — Scene 5: build up a big entrance

★★ a guest: In the first take of scene 5, Luca's entrance kept jumping between front and back views, long pauses slowed it down, and Lucy's reaction was lost because her face had turned away. — 00:13:54
★★★ a guest: For a character's entrance the presenter builds anticipation before the face is seen, starts the dialogue while the characters are still moving instead of a long silent walk, and moves the camera closer to show the reaction. — 00:13:54
★★★ a guest: When a take needs changes, the presenter points them out to ChatGPT, has it update the generation prompt, generates another version and compares the two. — 00:13:54 · ChatGPT
★★ a guest: Near a short drama's ending the presenter picks up the pace: changing angles, an entrance and small interactions make it feel that something big is about to happen. — 00:14:43
★★★ a guest: For a location used in only one shot (the dressing room), the presenter skips a separate location image: ChatGPT uses the first room as a style reference and the new space is described in the prompt. — 00:14:43 · ChatGPT

00:15:12 — Scene 6: end right before the reveal

★★★ a guest: The presenter ends a short-drama episode on a cliffhanger, just before the awaited answer (a decision, a reveal of who someone is, a new look), so viewers want the next episode. — 00:15:19

00:16:04 — The full episode again

★★ a guest: The presenter says a cinematic look gives short dramas a big advantage over sloppy AI videos, though those also get views and make money. — 00:18:02

00:18:05 — Step 5: three ways to monetise short dramas

★★★ a guest: The presenter lists three ways to earn from short dramas: make them for clients, take paid brand integrations, and sell access to the next episode. — 00:18:28
★★ a guest: The presenter says brands like to work with short-drama creators because viewers keep watching for the story while getting to know the product, which builds brand awareness and can drive sales. — 00:18:28
★★ a guest: To sell access to the next episode of a short drama from a new page, the presenter relies on free episodes that make people curious about what happens next. — 00:18:28
★★ a guest: The presenter suggests offering a short drama as a monthly subscription on a creator platform and testing a price such as $5 or $10. — 00:18:56
★★ a guest: In the presenter's example, 500 subscribers paying $5 a month would bring in $2,500 before fees and costs; the challenge is a series people keep paying for. (hypothetical) — 00:18:56
★★ a guest: The presenter invites viewers to join Higgsfield's Discord contest, held because the server reached 200,000 members, with a prize pool of 200,000 credits; the details are in the server. — 00:18:56 · Discord

00:19:25 — The Discord contest and sign-off

In the description and comments

★★ Higgsfield gives a full breakdown of how it made its viral AI short drama, which has passed 60 million views: the story, Soul 2.0 characters and Seedance 2.5 scenes with failed takes and fixes. — video description · Soul 2.0, Seedance 2.5
★★ Higgsfield says its short-drama skill keeps a character's voice consistent without a voice reference by writing one fixed voice description per character: accent, pitch, tone and pace. — creator comment
★ Higgsfield agrees with a viewer that the story is now the hard part of making an AI film. — creator comment
★ Higgsfield: Asked where the short drama can be seen, Higgsfield links to it as an Instagram reel. — creator comment

Threads in this video

restate · How do you connect a chat or coding agent to a generation platform? — The presenter connects Higgsfield to ChatGPT and builds the whole short-drama episode in one ChatGPT conversation. (00:02:21)
restate · How do you structure a short-form AI video for social media? — The presenter builds a short drama's story around four parts that every short drama is built on: the hook, the context, the escalation and the cliffhanger. (00:02:48)
restate · Which image model should be used for keyframes, edits, text and outfits? — The presenter uses Soul 2.0, called the best image model for ultra-realistic faces and locations, to avoid the typical AI face. (00:03:36)
restate · How do you keep a consistent colour palette and visual style across generated images? — The presenter describes Lucy to ChatGPT, has it generate options with Soul 2.0, and adds early-2000s lighting, colours and film texture to the description so she fits the story's world. (00:03:36)
restate · How do you keep a character the same across shots? — The presenter's rules for a character sheet: a plain grey background, three views (full body front, back, face close-up) and the face erased from the full-body front view, so the video model has one clear face reference instead of using different faces in different shots. (00:04:52)
first on record · How do you change one part of an image and leave the rest untouched? — For a character's second look the presenter has ChatGPT edit the character sheet with Seedream 5.0: hair down, glasses removed, face and outfit kept exactly the same. (00:05:46)
restate · How do you keep a location, product or prop the same across shots? — The presenter generates the casting-room location in a three-quarter view showing the table, the runway and the entrance, because with a straight-on view the video model must guess the rest of the room when angles change; showing more of the layout keeps the location consistent. (00:06:39)
restate · How do you write the script for an AI film? — The presenter gives ChatGPT the story outline and asks for a scene-by-scene script around the chosen cast and location, with dialogue, action and a generation prompt for each scene. (00:06:39)
restate · How do you make a hook that stops viewers scrolling? — The presenter opens a short drama's hook with the peak of the scene: the viewer sees the loudest moment without knowing how it came about and keeps watching to find out. (00:07:52)
first on record · How do you get a specific accent in generated speech? — To get an accent right in generated speech, the presenter describes how it sounds in detail; for Raphael's French accent he asked for a soft uvular R made at the back of the throat. (00:08:58)
restate · How do you keep a character's voice the same across shots? — The presenter gives each character a voice description, has ChatGPT carry it into every scene where that character speaks, and listens to the takes together to check the voices match. (00:10:11)
restate · How do you direct a character's acting in a video prompt? — In a Seedance 2.5 scene prompt the presenter describes both performances (Lucy apologises nervously, Raphael's boredom turns into curiosity) and spells out the movement: he gets up from the red chair, leaves the table and crosses the floor to her. (00:11:35)
restate · When a generation fails, should you generate more takes, fix the prompt or start over? — When a take needs changes, the presenter points them out to ChatGPT, has it update the generation prompt, generates another version and compares the two. (00:13:54)
restate · How do you choose the location image for a generated scene? — For a location used in only one shot (the dressing room), the presenter skips a separate location image: ChatGPT uses the first room as a style reference and the new space is described in the prompt. (00:14:43)
reminder · Higgsfield: enter the Higgsfield Discord 200K Contest by 14 October 2026 (2026) — The presenter invites viewers to join Higgsfield's Discord contest, held because the server reached 200,000 members, with a prize pool of 200,000 credits; the details are in the server. (00:18:56)
restate · How do you avoid the AI (plastic) look? — The presenter uses Soul 2.0, called the best image model for ultra-realistic faces and locations, to avoid the typical AI face. (00:03:36)

Mentioned

Central: Accountant and fashion director short drama · character sheet · ChatGPT · Lucy · Seedance 2.5 · Soul 2.0

Also: Projects: Higgsfield Discord 200K Contest

Organizations: Discord

Characters: Luca · Raphael · Rebecca

Products: GPT Image · Higgsfield ChatGPT Plugin · Seedance · Seedream 5.0

Also centred on Accountant and fashion director short drama: The Easiest Way to Make an AI Series in ChatGPT #AI #ChatGPT · This AI short drama got 60M views and doesn't look like AI

Replied to by

← Previous: The easiest way to make viral what if videos · Next → You don't have to quit your job to start with AI