GPT-6 ASTRA + SEEDANCE 2.5 : la révolution des UGC IA

Business-focused summary (UGC IA production playbook)

Goal / positioning

The video claims to teach an end-to-end process (A to Z) for producing high-performing AI-generated UGC (“UGC IA”) by:

• Replicating proven ad structures
• Systematizing production for ongoing output

---

Workflow / playbook (structured steps)

Step 1 — Find winning ads and reverse-engineer them (TrendTrack + LLM)

Use Trend Track (mentions “V3”) to:

• Find winning ads to replicate
• Download the ad video and download the script

Then:

• Feed the video/script into an LLM video analysis workflow to decompose why the ad worked
• The speaker references using an LLM (mentions “Kai AI key” and a “PMini” tool) for video analysis, presented as highly effective

Extracted ad template (4-part structure)

• Hook / opening (Action begins): Example described: opening a sachet, talking about taste, and presenting the product early
• Demonstration: Show the product (example: little gummies), emphasized as “super important” for production quality
• Argument / benefits: State brand benefits
• Short call-to-action (CTA): Close with a CTA

Outcome: After decomposition, the process is positioned as a way to reproduce UGC IA “indefinitely”—i.e., create a repeatable creative system.

---

Step 2 — Select (generate) the AI “portrait” and validate realism

Generate portraits where:

• Photos are entirely AI-generated (not real)

Criteria for selecting the final person

• Craft prompts for “ultra-human faces” (speaker references prompt engagement and tools like “Daylights” and “Visible Skin”)
• Generate unlimited photos using Clo
• Select a person that matches the brand/personal style

Quality checks (do not ship obvious AI artifacts)

Ensure:

• Skin must not look plasticized
• Look for realism signals like wrinkles and freckles
• Avoid deformities, e.g. no deformed fingers
• Avoid undesirable artifacts (mention of “varnish” implying glare/texture issues)
• Rule: “Do not place any products on the image.”

---

Step 3 — Animate the portrait (two options)

Option 1: Subscription route

Use a tool referred to as Xfield and Sens 2.5:

• Drag image 1 + image 2 plus the prompt
• Add “words, gestures and rules” to animate

Option 2: Alternative route

A method “partner and I built” where:

• You avoid paying the Xfield subscription
• You pay only “off your loans” (mechanism unclear in the source)

---

Prompt engineering requirements (6 components)

The speaker provides a concrete prompt structure with six required sections:

1. Format

   • Vertical: 9:16
   • Duration: 20 seconds

2. References

   • Image 1 as reference
   • Instruction like “control exact woman” to keep identity consistent

3. Scene + camera

   • Example: subject “looks through her phone”
   • Phone is filming and is out of frame (explicit camera composition instruction)

4. Timeline

   • Describe actions by moment (time-sliced instructions)

5. Continuity / connections (critical)

   • Keep consistent actions/objects across time
   • Example: same gummies in the right hand every time
   • Prevent “bugs” / glitches / plan breaks

6. Audio + constraints (critical)

   • Maintain one consistent adult female voice throughout
   • Purpose: prevent voice variation across segments

---

Step 4 — Turn it into an operating system (“internal workflow” / studio)

To scale output, the speaker emphasizes building a repeatable workflow.

System elements

• Your own internal studio (workflow environment)

Inputs include:

• Brief (target + concept/idea)
• Gesture and skills
• Products and personas
• Context
• Photo promises and video promises (prompts/templates)

Tools mentioned:

• Kai AI for linking portrait generation
  • Mentions GPT2 for portrait
  • Mentions Sidence for video

Capabilities claimed

• Keep validated sets/personas and repeatedly generate content
• Test a multitude of formats (creative iteration)

---

Metrics / KPIs / targets

• No explicit numeric business KPIs are provided (e.g., revenue, CAC, LTV, conversion rate, churn, growth targets)
• “Performance” is framed qualitatively:
  • Replicating “winning ads”
  • Production that can be “indefinitely” reproducible
  • Testing a multitude of formats

---

Concrete business examples / cases cited

• The speaker uses examples involving:
  • ADIA’s Brandy Ecom (as the working context)
  • A specific brand example: Gruns (used to demonstrate the 4-part ad structure with gummies and a sachet/taste hook)

---

Actionable recommendations (what to do next)

• Build a creative reverse-engineering loop
  1. Find winning ads in TrendTrack
  2. Download video + script
  3. Use an LLM to decompose structure
  4. Rebuild the same 4-part template (hook → demo → benefits → CTA)
• Enforce quality constraints in portrait and animation:
  • Check skin realism, fingers, artifacts
  • Maintain continuity (hand/object consistency)
  • Lock audio to one consistent voice
• Operationalize into a repeatable studio workflow:
  • Store personas/products/context
  • Reuse prompt templates (“photo promises” / “video promises”)
  • Rapidly test multiple formats

---

Presenters / sources mentioned

• Jeanlexandre (speaker)

Tools/platforms referenced in the process

• Trend Track (including “V3”)
• LLM / “PMini” (via Kai AI key mentioned)
• Clo
• Xfield
• Sens 2.5
• Kai AI (with mention of GPT2 for portrait)
• “Sidence” (video tool referenced)

Business examples/brands

• ADIA’s Brandy Ecom
• Gruns