Source: raw/5-Step_Workflow_-_How_to_Make_Viral_AI_Short_Dramas_60M_Views.md · product facts from ai-research/higgsfield-seedance-2-5-page-2026-11-11.md

Creator: Higgsfield AI (the vendor’s own YouTube channel) · URL: https://www.youtube.com/watch?v=cBxKZcalwkM · Platform: YouTube · Fetched: 2026-11-11 · Skill level: operator · Disclosure: first-party marketing. Every tool in the workflow is sold through Higgsfield, the video ends with a Higgsfield Discord contest, and the drama’s own plot has the director ask the heroine to “model for Soul”, Higgsfield’s image model.

Higgsfield’s channel walks through a 1-to-3-minute AI short drama, made entirely in one ChatGPT conversation with the Higgsfield plugin connected. The episode follows Lucy, an accountant at a fashion house, who trips into a casting and is picked as the house’s new face. The method has two halves. First, build named reference assets (character sheets and a location) with image models. Then write a scene-by-scene script on a four-part story structure and generate each scene with Seedance 2.5, fixing takes by telling ChatGPT exactly what to change. The channel says the original drama passed 60 million views; that figure is Higgsfield’s own claim.

Key Takeaways

  • Everything happens in one ChatGPT chat with the Higgsfield plugin connected. Keep working in that same chat so the references and names stay in context.
  • Each model does one job:
    • Soul 2: faces, group casts and locations. The presenter calls it “the best image model for ultra-realistic faces and locations”, to avoid “that typical AI face”.
    • GPT Image (version not stated; probably GPT Image 2 ^[inferred]): combines a face and an outfit into one character sheet, “because it works really well with references and preserves the details.”
    • Seedream 5.0: edits an existing sheet into a second look (hair down, no glasses) while keeping face and outfit.
    • Seedance 2.5: renders the scenes, with dialogue and voices.
  • Name every reference and reuse the names (LucyFace, LucyOutfit). Once a chat holds several characters and versions, names stop the model picking an older one.
  • Character sheet, three rules: plain grey background, not white or black; three views (full-body front, back, face close-up); and erase the face from the full-body front view, so the video model has “one clear face reference to follow”.
  • Story structure in four parts: the hook, the context, the escalation and the cliffhanger. Pick a familiar premise first: a Cinderella story, love against the odds, or a rivalry.
  • Fix takes by describing the change, not by re-rolling. Keep what works (he kept the acting and camera work), name what is wrong, and have ChatGPT rewrite the prompt. For an accent, describe the sound itself: for a French accent he asked for “a soft uvular R”.
  • Realism is the pitch. The presenter’s argument is that most AI short dramas on social media are “slop”, so footage that looks like a film clip stands out.

The workflow, step by step

1. Premise and length

  • Pick a situation viewers understand at once. His three starting points:
    • A Cinderella story: someone ordinary gets an unexpected opportunity. (His choice.)
    • Love against the odds: two people want to be together, but something is in the way.
    • A rivalry: two people want the same thing, and only one can get it.
  • Set the length: one to three minutes for a short drama.
  • Write the premise in a sentence. His: Lucy, an accountant at a fashion house, interrupts a casting, trips, and catches the director’s attention. He thinks she could become the face of the house.

2. Characters (Soul 2, then GPT Image, then Seedream 5.0)

  1. Face. Describe the character to ChatGPT and ask for a few options with Soul. Include the look of the world: he described “early-2000s lighting, colors, and film texture” so Lucy already fitted the story. Reject options where the face looks washed out.
  2. Name the pick (LucyFace) and tell ChatGPT to use that reference in the next image.
  3. Outfit. Describe the clothes and ask for an outfit reference (LucyOutfit). You can upload a reference image instead and say which details to keep.
  4. Character sheet. Ask GPT Image to combine LucyFace and LucyOutfit into one sheet, following the three rules above.
  5. Second look. He needed Lucy with her hair in a bun and glasses on, then with hair down and no glasses after the fall. Make the first sheet, then ask for it to be edited with Seedream 5.0: “Let her hair down, remove the glasses, and keep her face and outfit the exact same.”
  6. Small details. For jewellery or similar, generate a separate close-up with Soul and attach it with the face and outfit references when building the sheet. He did this for the director’s necklace.
  7. Crowds. For the casting models he had Soul generate the whole group in one image (“different faces, different poses, everyone dressed in black”) and used that image as the reference for the whole scene.
  8. Repeat for the rest of the cast: Raphael the director, Rebecca his assistant, and Luca the stylist.

3. Location

  • Describe the room by what has to happen in it. His casting room needed a table for Raphael, a clear path for the models, and space for Lucy to walk in.
  • Ask for a three-quarter view, not straight-on. A straight-on shot “might look great”, but when the camera angle changes the video model has to guess the rest of the room. Showing more of the layout keeps the location consistent from shot to shot.
  • For a one-shot location, skip the extra image. For the dressing-room shot he asked ChatGPT to use the casting room as a style reference and describe the new space in the prompt.

4. Script and prompts

  • Give ChatGPT the story outline and ask for a scene-by-scene script built around the chosen cast and location: dialogue, action and a generation prompt for each scene.
  • He attached two skills, linked in the video description (not in the transcript):
    • A script skill with “a hundred story ideas built in” that picks the strongest.
    • A prompt skill for camera, lighting and cinematic direction, “built specifically for Seedance, so your generation win rate goes way up.”

5. Generate, review, fix (Seedance 2.5)

Generate scene by scene and review each one before moving on. What each scene of his six had to do:

SceneStory partWhat he did or fixed
1Hook and contextOpen on “a peek” of the loudest moment so viewers stay to find out how it happened. Keep “emotional swings” coming every 20 to 30 seconds: “a starting point, then hope, then another hit, then a chance.” He fixed an unrealistic fall, added models waiting in the background so the casting felt busy, removed the assistant’s headshot from the model photos, and replaced an American accent with the uvular-R description.
2The turnLucy walks in, trips, and gets noticed. Build it on a feeling the audience recognises (“I don’t belong here, I just embarrassed myself, and now I’m gonna get in trouble”), then turn it into an opportunity.
3EscalationNew characters, growing conflict, things the viewer doesn’t expect. He described both performances (Lucy apologises nervously; Raphael’s boredom turns to curiosity) and spelled out the blocking: he “gets up from the red chair, leaves the table, and crosses the floor toward her.”
4EscalationRaphael makes the offer. A separate close-up of Rebecca calling Luca was generated as its own shot.
5EscalationFirst take: Luca’s entrance jumped between front and back, long pauses slowed it, and Lucy’s face turned away. Fix: build anticipation before his face is seen, start the dialogue while they are still moving, have Luca lift a strand of her hair as he introduces himself, and move the camera closer. He then compared the two takes.
6CliffhangerA 10-second scene: Luca brings her to the mirror, says they are ready to begin, cut to black. End “just as we’re about to get the answer”, such as a decision, a reveal or a new look.
  • Voices: give each character a written voice description and ask ChatGPT to carry it into every scene where they speak. Then listen to the takes together to check the voices match.

Monetisation (creator’s suggestions, not verified)

  • Make short dramas for clients.
  • Paid brand integrations: viewers stay for the story and get familiar with the product along the way.
  • Sell access to the next episode through a creator platform with a monthly subscription. He suggests testing 10, and works out that 500 subscribers at 2,500 a month “before fees and costs”. Free episodes are the funnel. He says the hard part is “building a series people want to keep paying for.”

Try It

  1. Connect the Higgsfield plugin to ChatGPT (setup in Higgsfield in ChatGPT with GPT-6 Astra) and open one chat for the whole episode.
  2. Write a one-sentence premise from one of the three starting points, and a four-line outline: hook, context, escalation, cliffhanger.
  3. Build one character to the three-rule sheet (grey background, three views, no face on the front full-body view) and name each reference. Make a second look by editing the sheet rather than starting again.
  4. Generate the location as a three-quarter view showing every area the scenes use.
  5. Generate scene 1 only. Before scene 2, list what to keep and what to fix, and have ChatGPT rewrite that scene’s prompt. For accents and voices, describe the sound.
  6. Check that a short pilot works before building the paywall idea: the creator’s own caveat is that people must want the next episode first.

Implementation

Tool/Service: ChatGPT with the Higgsfield plugin; Higgsfield models Soul 2, GPT Image, Seedream 5.0 and Seedance 2.5. Setup: Higgsfield MCP section → ChatGPT → Add Higgsfield plugin → install and sign in (see Higgsfield MCP). Two prompt skills from the video description, if you want his script and Seedance prompting rules. Cost: not stated. Generations use Higgsfield credits. Integration notes:

  • Seedance 2.5 on Higgsfield, per its product page: up to 30 seconds per generation, any aspect ratio from 9:16 to 21:9, audio generated in the same pass, up to 50 image and clip references, and region-level edits that repaint one area without regenerating the clip. Higgsfield’s page says “native 1080p”; another source in this wiki recorded 720p on a different provider in August 2026.
  • The page’s own prompting advice matches the video: “Write the prompt like a shot list: subject, camera, lighting, mood, sound,” and use region edits instead of re-rolling.
  • His scenes were generated one at a time; scene 6 is stated as 10 seconds. How long the other scenes were is not stated.

Open Questions

  • The two skills are not in the transcript. Their contents (the hundred story ideas, the Seedance prompt rules) are behind the description link and were not fetched.
  • Which GPT Image version combined the sheets is not stated.
  • Credits per episode and the number of takes per scene are not stated.
  • View counts (20 million in 24 hours; over 60 million total) are Higgsfield’s own figures for a drama it produced, and could not be checked.
  • Whether region editing was used for any fix is not said; the fixes shown were prompt rewrites and new generations.