Model

Generates a matched headshot, full-body casting digital, and 4-panel character reference sheet for a brand model character.

Three images are produced:

1. Headshot — tight casting portrait, straight-on, clinical flash lighting
2. Full body — full-length casting digital using the headshot as a face reference
3. Character sheet — 4-panel reference grid (front / left profile / right profile / rear), 16:9, designed to feed Kling's "Bind elements to enhance consistency" feature so the character's face stays locked across video generations downstream

All three saved to ./brands/[brand-name]/models/[model-name]/.

The character sheet is on by default. Skip with --skip-character-sheet if you don't plan to feed downstream video tools that need a multi-angle reference (saves ~$0.12).

Two modes:

• Default — invented character driven by text characteristics. Use when the user is creating a model from scratch.
• Reference — user supplies a photo of a real person. The headshot is generated via image-reference edit with an identity-lock prompt so the output actually looks like them. Fullbody then uses the new headshot as the face ref. Same cost. Use when the user wants the model to resemble a specific person (client photo, creator likeness, etc.).

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 to set up their brand

Step 2 — Ask for a reference image (optional)

Before collecting characteristics, ask: "Do you have a reference photo of the person you want this model to look like? If so, paste the path or drop it into the chat. Skip if you're inventing a new character from scratch."

• Reference provided → save it to the model folder as reference.png (or the original extension). The generator will use it in reference mode so the output actually resembles the person. Still collect characteristics in Step 3 — they guide wardrobe, expression, and anything the image alone can't convey.
• No reference → proceed as a fresh invented character.

Step 3 — Collect characteristics

Ask the user for all of the following in one grouped message. Tell them to skip anything that doesn't apply (e.g. facial hair for a female model).

I need a few details to build your model. Fill in what applies:

DEMOGRAPHICS
Age:
Sex:
Ethnicity:

PHYSICAL FEATURES
Skin tone:
Face shape:
Jawline:
Cheekbones:
Cheeks:
Eye shape:
Eye colour:
Eyebrows:
Nose:
Lips:
Freckles:
Facial hair:

HAIR
Colour:
Style:

BUILD
Height:
Body type:

AESTHETIC NOTES
Tone:        (e.g. Natural, editorial, high fashion, athletic, commercial)
Expression:  (leave blank for neutral — the default)

Confirm back the full characteristics list before proceeding. Ask if they want to change anything.

Step 4 — Name the model

Ask: "What would you like to name this model? This becomes the folder name — e.g. jade, marcus-02, hero-female."

Slug the name: lowercase, hyphens, no spaces or special characters.

Step 5 — Write model-spec.json

Create the folder and write the spec:

brands/[brand-name]/models/[model-name]/
  model-spec.json
  characteristics.md

model-spec.json format:

{
  "model_name": "[model-name]",
  "brand": "[brand-name]",
  "characteristics": "Age: 23\nSex: Female\nEthnicity: Asian\nSkin tone: Slightly olive, slight freckles\nFace shape: Oval\nJawline: Defined\nCheekbones: Prominent, high\nCheeks: Natural, strong and defined\nEye shape: Almond\nEye colour: Green\nEyebrows: Natural, medium thickness\nNose: Small, straight, lightly rounded tip\nLips: Plump, natural finish\nFreckles: Lightly scattered on nose and cheeks\nFacial hair: None\nHair colour: Light brown\nHair style: Mid-length, slight waves, centre part\nHeight: Average\nBody type: Slim\nTone: Natural, high fashion"
}

Format the characteristics as a flat newline-separated key: value list. Include only fields the user provided — omit blanks.

characteristics.md — a readable reference copy:

# [Model Name] — Characteristics

## Demographics
Age: 23
Sex: Female
Ethnicity: Asian

## Physical features
Skin tone: Slightly olive, slight freckles
Face shape: Oval
...

## Hair
Colour: Light brown
Style: Mid-length, slight waves, centre part

## Build
Height: Average
Body type: Slim

## Aesthetic
Tone: Natural, high fashion

Step 6 — Generate

Run from the project root.

Default (invented character, no reference photo):

python3 skills/references/generate-model.py brands/[brand-name]/models/[model-name]

1. Headshot — text-to-image at 2K, 3:4 (fal-ai/nano-banana-2)
2. Full body — image-reference edit at 2K, 3:4 (fal-ai/nano-banana-2/edit), using the headshot as face reference
3. Character sheet — image-reference edit at 2K, 16:9, using the headshot as face reference, prompted as a 4-panel grid (front / left profile / right profile / rear)

Reference mode (user provided a photo in Step 2):

python3 skills/references/generate-model.py brands/[brand-name]/models/[model-name] --reference brands/[brand-name]/models/[model-name]/reference.png

1. Headshot — image-reference edit at 2K, 3:4, using the raw reference photo as face reference, with an identity-lock prompt prefix that forces the output to preserve the subject's real face
2. Full body — image-reference edit at 2K, 3:4, using the new headshot as face reference (same identity-lock prefix)
3. Character sheet — image-reference edit at 2K, 16:9, using the new headshot as face reference (same identity-lock prefix)

Use reference mode when the output needs to actually resemble a real person (client photo, creator likeness, etc.). The characteristics text still guides wardrobe, expression, and anything the image alone can't convey.

Skip the character sheet (cost-conscious or you don't need a multi-angle reference for downstream video):

python3 skills/references/generate-model.py brands/[brand-name]/models/[model-name] --skip-character-sheet

Cost: ~$0.36 total with character sheet (3 × $0.12 at 2K) · ~$0.24 with --skip-character-sheet · same in either reference mode.

Step 7 — Present the results

Once generation completes:

• Tell the user the exact save paths, e.g.:
• brands/[brand-name]/models/[model-name]/headshot.png
• brands/[brand-name]/models/[model-name]/fullbody.png
• brands/[brand-name]/models/[model-name]/character-sheet.png (unless skipped)
• Ask: "Happy with this model? Or would you like to regenerate or tweak the characteristics?"

Once the user is happy, suggest the next step:

"Your model is ready. When you're ready to place clothing on them, run /style to dress the model in a product from your brand. The character-sheet.png is what you upload to Kling's 'Bind elements' feature in /add-motion (or directly on kling.ai) to lock the character's face across video generations."

Regeneration and iteration

Regenerate all images (same characteristics):

python3 skills/references/generate-model.py brands/[brand-name]/models/[model-name]

Add --reference path/to/photo.png if the original run used a reference image (and you still want likeness preserved). Overwrites headshot.png, fullbody.png, and character-sheet.png. Add --skip-character-sheet to only regenerate the first two.

Tweak characteristics and regenerate:

1. Ask which characteristics to change
2. Update model-spec.json and characteristics.md
3. Re-run the script

Regenerate full body or character sheet only (keep headshot): Not supported directly — the full body and character sheet both use the headshot as face reference, so all three regenerate together for a consistent match. The headshot step is fast (~15 seconds).

Notes

• Each brand has its own models/ subfolder — multiple brands can have separate model rosters in the same project
• Models within a brand can be reused across all styling and shoot workflows for that brand
• Multiple models per brand are fine: brands/[brand]/models/jade/, brands/[brand]/models/marcus/
• If the script fails with a FAL API error, check .env has a valid FAL_KEY