My AI UGC looks 100% real because I don't use Seedance (copy me)

Summary of the video (tech concepts + workflow features)

The video explains a workflow for creating hyper-realistic AI UGC avatars (the speaker references “Omni” as the tool/model) that looks more natural than typical AI UGC. The presenter claims this approach is better than “Seedance/Seed Dance” while also being cheaper, and focuses on tuning three key areas:

• Voice stability
• Image-to-image prompting (using real reference creators)
• Output quality control

---

The 3 most important tweaks for realism

1) Voice stability between takes (lock dialogue consistency)

The creator emphasizes that after setting up a first reference image, most of the time is spent iterating on the first generated video/take.

They recommend:

• Generate the same take multiple times (e.g., three times).
• When satisfied, regenerate to verify stability.

Key mechanism: “fixing the prompt”

• Keep everything the same except the dialogue text.
• Use the same reference image + same prompt, only changing what the avatar says.
• Result: the avatar maintains consistent energy, intonation, and accent across takes.

Example/testing described

• A reference image from TikTok was used to create multiple different characters using the same query.
• The system reportedly preserves consistent voice behavior and adds appropriate emphasis based on who is speaking on screen.

---

2) Image-to-image prompting (turn real UGC creators into avatars using references)

The presenter argues that realism depends heavily on image reference quality, because base image-generation models often miss real pores/imperfections/noise present in genuine footage.

How to choose good reference images

They provide practical criteria, especially:

• Framing

  • Cropping/frame choice is subjective, but keep the subject not too close to the camera to allow room for imperfections that make results look less “AI.”

• Background

  • Prefer neutral backgrounds that won’t distract.
  • Avoid references where specific author identifiers appear (e.g., a name on a wall).
  • Add visual interest only if it fits the scene (e.g., plants).

• Lighting (most important)

  • Strong emphasis: avoid overexposure.
  • Bad signs:
    • Overexposed highlights on nose/face
    • “Waxy/glassy” skin from bright references
    • Eye/head glare (“glint” in eyes or shiny forehead) that AI animates poorly
  • Preferred:
    • Dimmer, non-overexposed lighting
    • Choose references that are too dark rather than too light when forced to pick
  • The presenter claims even Omni struggles with harsh lighting, so lighting is the key constraint.

Warning against “deepfake” creation

They stress that deepfakes are illegal.

Instead of replicating someone exactly, they suggest genetic/identity changes, such as:

• eye color
• facial structure
• ethnicity

Otherwise, they warn about legal liability.

Why AI-written prompts can ruin results

They claim that when users ask an AI to write prompts, it tends to add extra instructions like:

• “make it realistic”
• “add pores”
• “natural lighting”

Presenter’s claim: these details are already captured in the reference image, so extra instructions can damage the output and make it look less natural.

Simple prompting approach demonstrated

Instead of “magic prompts,” they use very direct prompts that mainly specify the new avatar identity/appearance while keeping the reference consistent.

Example prompt content (described in subtitles):

• Change avatar to a big strong Norwegian man with specific accessories (knot on head, black T-shirt, Apple headphones, microphone, axe pendant).
• Change avatar to a young Latina woman (hair in bun, AirPods, gold necklace, rose gold watch).

Reported outcome

• Same pose/background/hand position preserved, but the avatar becomes a different genetic identity and animates more believably (e.g., skin texture and hair detail without overexposed lighting).

---

3) Ensuring top output quality (and keeping the recipe consistent)

The presenter explains why some viewers’ results don’t match the workflow output. Most issues trace back to:

1. Reference images not being selected correctly
2. Using different models or skipping the reference-image stage

They recommend:

• If you want the same results, follow a 1:1 guide:
  • Use the same process
  • Use the same reference images
  • Use the same models

They note that if you change the “recipe” (e.g., different models), you’ll get different outputs. They also mention they’re experimenting with other models, but caution that changing models changes the result.

---

Resources / guides mentioned

• The speaker says they will provide resources, images and prompts, and a cheat sheet (and likely a guide for the workflow).
• They reference “part one” and “part two” as earlier step-by-step explanations of the process.

---

Main speakers / sources

• Main speaker: The video’s presenter/author (creator of the workflow), referencing their tool/model “Omni” and comparing against “Seedance/Seed Dance.”
• Source images referenced: Users’ submissions and a TikTok reference image used as an example in the workflow.