Product Modelling Cloning Influencer

Product Modelling — Cloning an Influencer

Reference video / photos → character sheet → identity-locked stills → product videos

You clone a real person's likeness and use it to model a product across stills and videos, without distorting the face. Every visual claim is verified against frames you actually extracted and viewed (see references/ and the companion grid-review skill).

Tooling is zero-install: Python with imageio, imageio_ffmpeg (bundles its own ffmpeg — no system ffmpeg/Homebrew), and Pillow. Video/image generation runs through Higgsfield MCP.

---

THE GOLDEN RULE — Do Not Distort The Face

The person is real. Their likeness is the product's credibility. Every generation step must anchor to real reference frames and say so in the prompt (identity IDENTICAL to reference, do not distort the face). When you can avoid AI-inventing a face at all (montage from real frames, local edit), prefer it. Verify identity on every output.

---

STEP 0 — Setup (one-time, no admin)

python3 -c "import imageio, imageio_ffmpeg, PIL" 2>/dev/null && echo OK \
  || pip3 install --user imageio imageio-ffmpeg pillow

imageio_ffmpeg.get_ffmpeg_exe() returns a full ffmpeg binary (libx264 included) even when the ffmpeg CLI is absent. Use it for ALL extraction, stitching, and encoding.

---

STEP 1 — Extract & pick the hero frame

Extract frames from the reference video, then pick the clearest frontal, three-quarter, and profile of the face (and full-body shots for wardrobe).

import imageio_ffmpeg, subprocess
FF = imageio_ffmpeg.get_ffmpeg_exe()
subprocess.run([FF,"-i","ref.mp4","-start_number","1","frames/frame_%05d.png"])

Then view a spread of frames (Read tool) and choose: large, sharp, near-frontal, both eyes visible, neutral expression. Note the frame numbers — they become your canonical anchors.

---

STEP 2 — Build the character sheet

Two deliverables:

1. Written spec (references/character-sheet-template.md) — face shape, skin, eyes, brows,

nose, lips, cheeks, hair, makeup, wardrobe, eyewear, expression + a face-lock prompt fragment

• a canonical reference frames table (which frame = which angle).

1. Visual montage sheet — a labelled contact sheet built from the REAL frames (zero distortion).

Use scripts/build_sheet.py as the template (crops frontal/3-4/profile + full-body looks, draws labels + a spec panel).

This written spec is reused as the prompt backbone for every later generation.

---

STEP 3 — Generate ultrarealistic identity-locked stills

Higgsfield generate_image. Upload the real reference frames first (see uploads in references/gotchas-and-uploads.md) and pass them as medias[].role="image".

• Model: nano_banana_pro (Google "Nano Banana Pro"). NOTE: the platform routes Pro requests

to engine nano_banana_2 in the job record — output is genuine 4K, this is expected. soul_2 is the alternative purpose-built character model.

• Resolution 4k, aspect 16:9 for sheets / 4:5/9:16 for portraits.
• Generate, in order: face sheet (front / 3-4L / 3-4R / profile / up-tilt / down),

full-body turnaround (front / 3-4 / side / back), then variants (no-glasses, alternate wardrobe, product variant e.g. clear vs tinted lens).

• Every prompt ends with: *"Keep identity and facial geometry identical to the reference; do not

distort the face."* Reuse the written face spec verbatim.

Product-variant trick: to change one attribute (e.g. tinted → clear lens) keep the frame refs, keep identity language, and state the change emphatically: "the SAME frame BUT fully clear transparent colorless lenses, no tint."

---

STEP 4 — Marketing Studio · Pro Virtual Try-On (one-click ad)

show_marketing_studio presets include Pro Virtual Try On (slug: virtual_try_on). Flow:

1. show_marketing_studio action=create type=avatar from a clean image of the person.
2. show_marketing_studio action=create type=product from the product shots.
3. `generatevideo model=marketingstudiovideo mode=virtualtryon aspectratio=9:16

resolution=1080p generateaudio=true productids=[...] avatars=[{id,type:'custom'}]`.

KEY LEARNING — wardrobe is driven by the AVATAR image, not the prompt. MS auto-rewrites your prompt (you'll see enhanced_prompt) and will invent clothing if the avatar has none. To control wardrobe (e.g. "put her in the brown blazer"), create the avatar from an image where she already wears it. To replicate an approved video's motion in new clothes: reuse the same prompt + product, swap to a wardrobe-correct avatar.

---

STEP 5 — Cinema Studio · multi-shot film

cinematic_studio_3_0 (most advanced) or cinematic_studio_video_v2 (genre control). Image-to-video, one clip per storyboard shot from your hero stills (medias[].role="start_image"), 9:16, dur 4–15. No audio param (silent — add music in edit). Then stitch locally (Step 7). Build a director grid sheet first to plan shots — see scripts/build_director_grid.py (storyboard cells from real stills + camera direction + Director's Note, in the seedance format).

---

STEP 6 — Posing videos · Kling 3.0

For active body posing + finger-level hand gestures, kling3_0 is the best engine (mode: std/pro/4k, sound on/off, dur 3–15, medias[].role="start_image"). Start from the full-body still for body poses, prompt a mix framing (open full-body → push to 3/4 for hand-to-face/adjust-glasses gestures). Use the posing-phrase library (references/posing-prompt-library.md). For "fast-paced": prompt rapid succession of poses, snappy rhythm, fast push-in/whip-pan (Kling renders continuous motion, not hard cuts — for true hard cuts, edit locally in Step 7).

---

STEP 7 — Local fast-cut montage / editing

Hard-cut editing, punch-in "zoom" reframes, and concatenation are done locally with the bundled ffmpeg — zero credits, full control. scripts/build_reel.py is the template: define a list of (clip, start, zoom, face_bias) cuts (~0.5s each), crop+scale each to 1080×1920, concat. This is also the reliable fallback when a single-gen montage engine refuses (e.g. Seedance NSFW block): generate clean takes, then cut down to a 20-shot reel locally.

---

STEP 8 — QA every output (always)

Use the grid-review skill: sample ~every 0.5s into a labelled contact sheet, view it, and run the failure checklist — identity drift, wardrobe drift, lens tint creeping back, warped product, hand/finger artifacts, text legibility. State which file you inspected. Remind the user you can only judge stills — they must playback-check motion + audio.

---

ENGINES, GOTCHAS, LIBRARIES — read the references

• references/engine-catalog.md — Higgsfield image+video model IDs, params, when to use each.
• references/gotchas-and-uploads.md — the upload flow, the Expect:100-continue stall fix,

NSFW false-positives, preset-recommendation notices, get_cost, "GPT 2.0 doesn't exist".

• references/posing-prompt-library.md — researched fashion posing + hand-gesture phrases.
• references/character-sheet-template.md — the written character-sheet format + face-lock fragment.
• scripts/ — build_sheet.py, build_director_grid.py, build_reel.py (reusable templates).

---

DECISION FLOW

Have a real person to clone?
  → Extract frames → pick hero → character sheet → identity-locked stills (always first)
Goal = quick branded try-on ad?        → Marketing Studio Pro VTO (avatar drives wardrobe)
Goal = cinematic multi-shot product film? → Cinema Studio 3.0 per-shot + local stitch
Goal = model posing / hand motion?      → Kling 3.0 (4k), posing-phrase library
Goal = punchy fast-cut reel?            → local ffmpeg montage from clean takes
Always: QA with grid-review. Never distort the face. Show plan + grid sheet before burning credits.

OPERATING PRINCIPLES (learned on the job)

1. Show the plan + a creative grid sheet BEFORE spending credits. Get a go/scope decision.
2. Never distort the face — anchor every gen to real frames; prefer real-frame edits when possible.
3. Verify, don't assume — grid every render; report what you actually saw, flag motion as unchecked.
4. Pick the engine for the job — try-on→Marketing Studio, posing→Kling, film→Cinema, edit→local.
5. Be honest about failures — NSFW blocks, wardrobe overrides, model routing — surface them.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: manvendrakumar-hub
• Source: manvendrakumar-hub/product-modelling-cloning-influencer
• License: MIT

Install and usage instructions live in the source repository linked above.