UGC Multiplier

Two-stage workflow:

1. Model swap — replace the face/model in the original UGC still with a new model (Nano Banana 2)
2. Motion transfer — apply the original UGC video's motion to the swapped image (Kling Motion Control v2.6 Standard)

Output saved to ./brands/[brand-name]/multiply-ugc/[output-name]/.

Step 1 — Select brand

Scan ./brands/ for subfolders that contain a brand-identity/visual-guidelines.md file.

• One brand found: use it automatically and confirm: "Using brand: [name]"
• Multiple brands found: list them and ask which to use
• None found: tell the user to run /brand first

Step 2 — Source UGC

Ask:

"Is the UGC you want to multiply one you created with the /ugc or /product-ugc skill, or an external video you've shot or downloaded?"

Option A — From `/ugc` skill output:

List available UGC outputs from ./brands/[brand-name]/ugc/:

Available UGC outputs:
  • jade-park-run
  • marcus-kitchen-hydration

The user selects one. Locate:

• Video: the latest .mp4 in brands/[brand-name]/ugc/[output-name]/
• Still image: derive the version number from the video filename (e.g. jade-park-run_v2-run1.mp4 → use jade-park-run_v2.png). This must be the image that was used to generate the video, not necessarily the latest image version.

If no video is found in the folder, tell the user: "No video found for [output-name] — make sure you generated the video in the /ugc skill first."

Option B — From `/product-ugc` skill output:

List available product UGC outputs from ./brands/[brand-name]/product-ugc/:

Available product UGC outputs:
  • jade-electrolytes-kitchen
  • marcus-protein-outdoors

The user selects one. Locate:

• Video: the latest .mp4 in brands/[brand-name]/product-ugc/[output-name]/
• Still image: derive the version number from the video filename (e.g. luna-energy-gel-yarra-bench_v1-run1.mp4 → use luna-energy-gel-yarra-bench_v1.png). This must be the image that was used to generate the video, not necessarily the latest image version.

If no video is found in the folder, tell the user: "No video found for [output-name] — make sure you generated the video in the /product-ugc skill first."

Option C — External UGC:

Create the uploads folder and ask the user to drop their video in:

mkdir -p brands/[brand-name]/multiply-ugc/uploads

"Drop your UGC video into: brands/[brand-name]/multiply-ugc/uploads/ Name it descriptively — e.g. jade-outdoor.mp4. Let me know the filename when it's in."

Wait for the user to confirm. Then:

• Video: brands/[brand-name]/multiply-ugc/uploads/[filename].mp4
• Still image: extract the first frame using ffmpeg:

  ffmpeg -i brands/[brand-name]/multiply-ugc/uploads/[filename].mp4 -vframes 1 -q:v 2 brands/[brand-name]/multiply-ugc/uploads/[filename]-frame1.jpg

Step 3 — Select target model

List available models from ./brands/[brand-name]/models/:

Available models:
  • jade
  • marcus-02
  • priya

The user selects one. The model's headshot will be used: brands/[brand-name]/models/[name]/headshot.png.

Confirm:

"Swapping into model: [model name] Headshot: brands/[brand-name]/models/[name]/headshot.png"

Step 4 — Name the output

Derive a slug from the source UGC name and the target model:

• jade-park-run-marcus
• outdoor-priya

Confirm with the user or let them rename.

Step 5 — Write multiply-ugc-spec.json

Create the output folder and write the spec:

brands/[brand-name]/multiply-ugc/[output-name]/
  multiply-ugc-spec.json

{
  "output_name": "jade-park-run-marcus",
  "brand": "[brand-name]",
  "source_ugc_image": "brands/[brand-name]/ugc/jade-park-run/jade-park-run_v2.png",
  "source_ugc_video": "brands/[brand-name]/ugc/jade-park-run/jade-park-run_v2-run1.mp4",
  "model_dir": "brands/[brand-name]/models/marcus"
}

For external UGC:

{
  "output_name": "outdoor-priya",
  "brand": "[brand-name]",
  "source_ugc_image": "brands/[brand-name]/multiply-ugc/uploads/outdoor-frame1.jpg",
  "source_ugc_video": "brands/[brand-name]/multiply-ugc/uploads/outdoor.mp4",
  "model_dir": "brands/[brand-name]/models/priya"
}

Step 6 — Generate

Run from the project root:

python3 skills/references/generate-multiply-ugc.py brands/[brand-name]/multiply-ugc/[output-name]

This runs both stages sequentially:

1. Uploads source_ugc_image (Image 1) and model headshot (Image 2) to Nano Banana 2 — prompt: "swap the model in image 1 with the model in image 2"
2. Saves swapped image as [output-name]-swap_v{n}.png
3. Uploads swapped image + source_ugc_video to Kling Motion Control v2.6 Standard with character_orientation: "video", no prompt
4. Downloads and saves the video as [output-name]_v{n}.mp4

Cost: ~$0.12 (model swap) + ~$0.07/sec of video output (Kling Motion Control)

Step 6.5 — Internal swap quality check (hidden)

After the swap image is generated, read and assess it internally before proceeding to motion transfer. Do not mention this check to the user.

Check for:

• Full model swap — the new model's face, hair, and overall appearance should be visible, not a face-only replacement
• Scene integrity — background, clothing, lighting, and environment from the source image should be preserved
• Colour accuracy — skin tone and hair colour should match the model's headshot, not be heavily tinted by scene lighting (e.g. orange cast, oversaturation)
• No split-image output — the result should be a single composited image, not a side-by-side comparison

If the check fails: silently re-run --swap-only once using the same spec. Do not tell the user.

If the re-run also fails: show the user the image and explain what's off. Ask if they'd like to regenerate again or adjust the source image.

If the check passes (first or second run): proceed directly to motion transfer without mentioning the check.

Step 7 — Present results

Tell the user the exact paths:

Generated:
  ✓ brands/[brand-name]/multiply-ugc/[output-name]/[output-name]-swap_v1.png
  ✓ brands/[brand-name]/multiply-ugc/[output-name]/[output-name]_v1.mp4

Ask: "Happy with these? You can:

• Regenerate — run again for another take
• Try a different model — update the spec with a new model_dir and regenerate"

To regenerate (both stages):

python3 skills/references/generate-multiply-ugc.py brands/[brand-name]/multiply-ugc/[output-name]

To run swap only (check the image before committing to motion transfer):

python3 skills/references/generate-multiply-ugc.py brands/[brand-name]/multiply-ugc/[output-name] --swap-only

To retry motion transfer only (reuses the latest swap image — useful after approving the swap or if Stage 2 timed out):

python3 skills/references/generate-multiply-ugc.py brands/[brand-name]/multiply-ugc/[output-name] --motion-only

Each run auto-increments — previous outputs are never overwritten.

Notes

• The still image drives the face and identity in the output video — a sharp, well-lit headshot in the source UGC produces better results
• character_orientation: "video" is always used — it supports up to 30s and handles complex motion better than "image" mode
• No prompt is passed to Kling — the motion is transferred purely from the reference video
• keep_original_sound: true is set by default — the original audio carries through to the output video
• Each brand has its own multiply-ugc/ subfolder — multiple brands can run independent workflows in the same project