3D-style image generation: 3D characters, product renders, isometric dioramas, 3D icons, 3D text, interior design renders, architectural visualization, 3D scenes, game assets. Use when generating 3D-style 2D images from text descriptions or reference photos (e.g. 3D character design, isometric diorama, 3D product render, interior design visualization, architectural render, 3D app icon, 3D text effect, game asset render).
Detected risks:
image-3d generates 3D-style images from text descriptions or reference photos, covering 3D characters, product renders, isometric dioramas, 3D icons, 3D text effects, interior design visualization, architectural rendering, 3D scenes, and game asset renders. It is built for the Starchild platform and solves the problem of producing a polished 3D-rendered look with granular category and style control. An important distinction the skill makes explicit: it produces 3D-style 2D images (PNG/JPG) that look 3D, not actual 3D model files (.glb, .obj, .fbx, .usdz) — for real 3D models, users need dedicated modeling services like Meshy.
The core principle is to call the provided script rather than re-implement proxy or billing plumbing. When invoked through the bash tool, the skill instructs using a python3 heredoc (python3 - <<'EOF') with from exports import generate_3d, because exec(open(...)) fails under /bin/bash -c and inside python3 -c (NameError on __file__). The generate_3d function accepts prompt, reference_path, reference_url, category (default "character"), style (default "default"), model ("nanopro" fast ~25s or "gpt" best quality ~150s), count (1–4), and aspect_ratio (1:1, 3:4, 4:3, 9:16, 16:9, auto-selected by category). A prompt is required for text-to-image mode unless a non-default style is specified. Categories include character (chibi, realistic, cartoon, fantasy), product (floating, exploded, turntable), and diorama (isometric, lowpoly, realistic), among others, each with prompt tips and auto aspect ratios (3:4 for characters, 1:1 for products/dioramas/icons/game assets, 16:9 for text/interiors/architecture/scenes).
It targets Starchild users creating 3D-styled visuals for characters, products, scenes, icons, and game assets. Results return a dict with a local_path per image; the skill stresses never handing users the raw fal.media URL (restrictive CSP headers) and instead delivering the already-downloaded local file from output/images/, embedding inline on Web or sending via send_to_telegram / send_to_wechat on those channels. For general creative images, product e-commerce photos, image editing, or identity-preserving portraits, the skill points to sibling skills image-create, image-ecommerce, image-edit, and image-portrait.
No. It generates 3D-style 2D images (PNG/JPG) that look 3D. It does not produce 3D model files (.glb, .obj, .fbx, .usdz); for those, users need dedicated 3D modeling services such as Meshy.
Use a python3 heredoc: python3 - <<'EOF', insert skills/image-3d into sys.path, then from exports import generate_3d. Pasting exec(open(...)) directly into bash fails, and exec inside python3 -c fails with NameError on __file__.
prompt, reference_path, reference_url, category (default "character"), style (default "default"), model ("nanopro" ~25s or "gpt" ~150s), count (1–4), and aspect_ratio (1:1, 3:4, 4:3, 9:16, 16:9). A prompt is required unless a non-default style is set, which provides a built-in template.
Never hand over the raw fal.media URL because fal serves files with restrictive CSP headers. Use each image's downloaded local_path under output/images/ — embed inline on Web, or send via send_to_telegram or send_to_wechat on those channels.
It is auto-selected by category: 3:4 for characters, 1:1 for products, dioramas, icons, and game assets, and 16:9 for text, interiors, architecture, and scenes.
Quick Setup:
.claude/skills/Repository
starchild-ai-agent/official-skills