image-portrait

Use this skill for all identity-consistent portrait generation requests on Starchild.

Covers: professional headshots, dating/social photos, artistic style transfers, themed/holiday portraits, photo series, digital avatars, children/family photos, ID/passport photos.

Core principle: call the provided script. Do not re-implement proxy/billing plumbing.

1. Quick start — single portrait (most common)

⚠️ Execution context — read this first. The code blocks below are Python, not shell commands. Starchild's bash tool runs /bin/bash -c, which cannot parse exec(open(...)) — pasting them directly into a bash command will fail with syntax error near unexpected token 'open'. Also, exec(open(...)) inside python3 -c fails with NameError: __file__ because the script uses __file__ for path resolution.

Use python3 - <<'EOF' with from exports import when calling via the bash tool:

python3 - <<'EOF'
import sys
sys.path.insert(0, "skills/image-portrait")
from exports import generate_portrait
result = generate_portrait(
    image_path="path/to/user/photo.jpg",
    style="professional",
)
print(result)
EOF

Copy

The heredoc (<<'EOF') preserves all quotes and newlines — no escaping needed.

exec(open('skills/image-portrait/generate_portrait.py').read())
result = generate_portrait(
    image_path="path/to/user/photo.jpg",
    style="professional",
)
# result -> {"success": True, "images": [{"local_path": "output/images/..."}], ...}

The script reads the local file, base64-encodes it, and sends it to fal.ai as a data URI — no manual URL publishing needed.

2. Quick start — public URL

exec(open('skills/image-portrait/generate_portrait.py').read())
result = generate_portrait(
    face_image_url="https://example.com/photo.jpg",
    style="anime",
)

3. Quick start — text-to-image (no reference photo)

exec(open('skills/image-portrait/generate_portrait.py').read())
result = generate_portrait(
    prompt="a young woman in cyberpunk armor, neon city background, rain",
    model="nanopro",
)

When no image_path or face_image_url is provided, the script uses the text-to-image endpoint (no /edit suffix).

Delivering the result to the user — IMPORTANT

Never hand the user the raw fal.media URL. fal serves files with restrictive CSP headers. The only reliable delivery path is the already-downloaded local file:

1. Use each image's local_path (e.g. output/images/xxx.png) — the script always downloads on success.
2. Tell the user the files are saved to output/images/ and viewable in the workspace file panel.
3. On Web channel, embed inline so the user can preview in chat:

   ![photo](output/images/<filename>.png)
   

   Copy
4. On Telegram / WeChat: send via send_to_telegram(file_path="output/images/...", message_type="image") or send_to_wechat(file_path="output/images/...", message_type="image").

4. Parameters

Image input rules:

• Provide image_path OR face_image_url for identity-consistent generation (edit mode).
• If both are given, image_path takes priority.
• Omit both for pure text-to-image generation (generate mode).

Prompt priority: prompt > style + scene > style > default (professional).

5. Style presets

A: Identity-consistent character styles

B: Personal showcase / dating / social

D: Themed / scene portraits

O: Digital avatar

T: Children & family

U: ID / passport photos

6. Model selection guide

Decision rules:

1. Default: always use nanopro unless the user explicitly requests otherwise.
2. Use nano2 when: user wants fastest results, is iterating on styles, generating many images, or says "quick", "draft", "fast".
3. Use gpt when: user says "highest quality", "best quality", "premium", or the scene is very complex with many specific details.

# Default (fast)
result = generate_portrait(image_path="photo.jpg", style="anime")

# High quality (user requested)
result = generate_portrait(image_path="photo.jpg", style="anime", model="gpt")

7. Custom scene examples

# Style + custom scene
result = generate_portrait(
    image_path="uploads/my_photo.jpg",
    style="professional",
    scene="in a modern office with city skyline view",
)

# Custom scene only (defaults to professional style base)
result = generate_portrait(
    image_path="uploads/my_photo.jpg",
    scene="standing on a beach at sunset, golden hour lighting",
)

# Fully custom prompt (overrides everything)
result = generate_portrait(
    image_path="uploads/my_photo.jpg",
    prompt="portrait of a person as a medieval knight, full plate armor, castle background, dramatic lighting, oil painting style",
)

# Different aspect ratio
result = generate_portrait(
    image_path="uploads/my_photo.jpg",
    style="cyberpunk",
    aspect_ratio="9:16",
)

# Multiple images
result = generate_portrait(
    image_path="uploads/my_photo.jpg",
    style="dating_cafe",
    count=4,
)

8. Prompt engineering best practices

When the user's request doesn't match any preset style, or when you need to construct a custom prompt, follow these guidelines (derived from reference skills: ai-headshot-generation, ai-avatar-generation, style-transfer, portrait-enhancement, character-design-sheet, avatar-portrait, nano-banana-pro, pet-portrait-generation).

Automatic likeness preservation

When a reference image is provided (edit mode), the script automatically prepends a likeness preservation instruction to every prompt. This ensures the generated portrait preserves the subject's facial identity. You do NOT need to add likeness instructions manually — the script handles it.

Exception: avatar styles (avatar_3d, avatar_gaming, avatar_vtuber) skip the likeness prefix because stylization takes priority over photographic likeness.

The 7-element prompt structure

Every effective portrait prompt should include these elements (from nano-banana-pro skill):

[subject], [outfit/attire], [pose/action], [expression], [background/setting], [lighting], [style/quality modifiers]

Key principles

1. Likeness vs. style balance (from avatar-portrait skill):

   • Too photorealistic = ignores requested style
   • Too stylized = loses resemblance to source person
   • For stylized portraits: emphasize "stylized but maintains individual features"
   • For photorealistic: emphasize "keep facial features recognizable"

2. Lighting is critical — always specify lighting type:

   • Studio: "soft diffused studio lighting", "Rembrandt chiaroscuro lighting"
   • Natural: "golden hour warm light", "dappled sunlight through trees"
   • Dramatic: "dramatic rim lighting", "volumetric light beams", "neon glow"
   • Flat: "even flat lighting with no shadows" (for ID photos)

3. Background specificity — vague backgrounds produce poor results:

   • ❌ "nice background"
   • ✅ "blurred modern office with glass windows and city view"
   • ✅ "clean neutral gray gradient studio background"
   • ✅ "background style should match the character style" (for avatars)

4. Lens/camera hints — help the model understand framing:

   • "85mm lens look, shallow depth of field" (portrait)
   • "head and shoulders framing" (headshot)
   • "full body, clean white background" (character design)
   • "close-up face, portrait orientation" (expression/avatar)

5. Quality anchors — add style quality references:

   • "professional photography quality", "magazine cover quality"
   • "National Geographic photography style" (adventure)
   • "League of Legends splash art style" (gaming)
   • "Pixar and Disney animation style" (3D avatar)
   • "Studio Ghibli inspired" (anime)
   • "fine art watercolor painting look" (watercolor)

6. Texture and material — for artistic styles, specify medium:

   • "visible impasto brushstrokes, canvas texture" (oil painting)
   • "loose expressive watercolor style, soft edges, beautiful color bleeds and washes" (watercolor)
   • "natural film grain, Kodak Portra emulation" (vintage)
   • "cel-shaded, clean line art, bold outlines" (anime)
   • "visible pixels but NOT a pixelated photo filter" (pixel art)

7. Expression guidance — be specific about mood:

   • ❌ "smiling"
   • ✅ "warm genuine smile, confident approachable expression"
   • ✅ "neutral calm expression with mouth closed" (ID photo)
   • ✅ "passionate expression, energetic" (musician)

Example: building a custom prompt

User: "I want a photo of me as a wizard in a magical forest"

result = generate_portrait(
    image_path="uploads/photo.jpg",
    prompt=(
        "fantasy wizard portrait, wearing mystical purple robes with glowing runes, "
        "ancient wooden staff with crystal orb, wise powerful expression, "
        "enchanted forest background with bioluminescent plants and floating particles, "
        "dramatic magical lighting with ethereal glow, "
        "high fantasy art style, detailed digital painting quality"
    ),
)
# Note: likeness prefix is auto-added because image_path is provided

Example: pixel art avatar (from avatar-portrait skill)

User: "Make me a retro pixel art avatar"

result = generate_portrait(
    image_path="uploads/photo.jpg",
    prompt=(
        "retro 16-bit pixel art portrait, visible pixels with clean lines, "
        "rich colors, consistent shading, stylized but maintains individual features, "
        "warm sunset cityscape background in matching pixel art style, "
        "head and shoulders, square format"
    ),
)

9. Photo series

Generate a coordinated set of themed portraits in one call. Pass a custom list of styles/scenes — the agent assembles the list based on the user's request.

exec(open('skills/image-portrait/generate_portrait.py').read())
result = generate_series(
    image_path="uploads/my_photo.jpg",
    series=[
        {"style": "professional"},
        {"style": "casual", "scene": "at a rooftop bar, sunset"},
        {"style": "anime"},
        {"prompt": "portrait as a superhero, cape flowing, city skyline"},
    ],
)
# result -> {"success": True, "images": [...4 images...], "series": "custom"}

Each item in the list is a dict with optional keys:

• style — any style key from §7 (e.g. "professional", "anime", "cyberpunk")
• scene — override the scene description (combined with the style template)
• prompt — fully custom prompt (ignores style/scene)

10. Intent recognition guide

Use this table to map user requests to the correct style/parameters:

When NOT to use this skill (routing)

This skill's core contract is identity preservation: whenever a reference photo is provided, a likeness prefix ("preserve the subject's exact facial features…") is prepended to every prompt (except the 3 avatar_* styles). This means:

• User wants to drastically change the face/identity or fully re-imagine the person (e.g. "make me look like a different person", heavy character redesign) → route to image-create (text-to-image) instead. The likeness prefix will fight the stylization and iterations won't converge.
• User wants strong stylization but still recognizable → stay here; use anime / avatar_3d etc.
• User wants to edit a non-person photo → image-edit.

If a request keeps failing to move away from the reference photo's look after 2+ iterations, that's the likeness contract working as designed — switch skills rather than re-prompting.

11. Provided scripts

12. Local testing

Set FAL_KEY env var to call fal.ai directly (bypasses sc-proxy):

# Single portrait
FAL_KEY=your-fal-key python3 skills/image-portrait/generate_portrait.py photo.jpg anime 1 nanopro

# Args: <image_path_or_url> [style] [count] [model]

13. Troubleshooting

14. Infrastructure (reference)

• Caller → sc-proxy → queue.fal.run/{model} → fal model providers
• All requests must include Authorization: Key fake-falai-key-12345 (proxy injects the real FAL_KEY)
• Pre-charge happens at submit. Poll/result calls are free.
• Local files are base64-encoded as data URIs — no separate upload step needed.
• Final images live at https://*.fal.media/... — public CDN, no auth needed for download.
• Cost tracking via _cost_track.py — records X-Credits-Used from sc-proxy response headers.

Model endpoints