version: "1.0.1" name: fal-ai-image description: "Generate/edit images via fal.ai. Supports Google Nano Banana Pro and OpenAI GPT Image 2 selected from config/.env. Supports reference images and strong text rendering. ALWAYS read SKILL.md before first use."

fal-ai-image

Generate images via fal.ai. The skill now supports two fal-hosted models:

• Google Nano Banana Pro (fal-ai/nano-banana-pro) — default, backward-compatible
• OpenAI GPT Image 2 (openai/gpt-image-2) — enabled from config/.env

Synonyms the agent should treat as equivalent:

• gpt = openai = GPT Image 2
• nano banana = google = gemini = Nano Banana Pro

Best for: infographics, text rendering, banners, photo edits, reference-based compositions.

STOP — Read Before Acting

• DO NOT use Pillow, ImageMagick, or post-processing for text/logo overlay unless the user explicitly asked for that workflow
• DO NOT use Generate mode when the user provided reference images — use Edit mode
• DO NOT assume GPT is active — check config/README.md logic; if no selector is configured, the skill stays on Nano Banana
• DO NOT guess provider-specific params — Nano Banana and GPT Image 2 use different schemas
• DO pass --model ... when the user explicitly asks for a specific model/provider or asks to compare providers
• DO NOT skip uploading local files — run upload.sh first to get URLs for edit.sh

Quick Start Decision

Reference images provided?  -> Edit mode   (upload.sh -> edit.sh)
Text-only generation?       -> Generate mode (generate.sh)
 
Model comes from config/.env:
  no selector set           -> Nano Banana Pro
  FAL_IMAGE_PROVIDER=openai -> GPT Image 2
  FAL_IMAGE_MODEL=...       -> exact override

Config

Requires FAL_KEY in config/.env or the environment.

Model selection:

1. --model — one-off override for the current command
2. FAL_IMAGE_MODEL — exact override in config
3. FAL_IMAGE_PROVIDER — google or openai
4. nothing set — default to Nano Banana Pro

Use --model whenever the user explicitly says things like:

• "сделай через GPT"
• "используй OpenAI"
• "сделай через Google / Gemini / Nano Banana"
• "какие тут есть провайдеры?"

Answer that the skill supports two choices and map the command like this:

• --model gpt for OpenAI GPT Image 2
• --model gemini or --model nano-banana for Nano Banana

OpenAI quality default:

• FAL_IMAGE_OPENAI_QUALITY=medium unless overridden with --quality
• this skill intentionally uses medium as the default for GPT to avoid expensive exploratory runs

Full setup and troubleshooting: config/README.md.

Read references only when needed:

• references/MODELS.md — selector precedence, aliases, cost heuristics
• references/EDITING.md — generate vs edit, mask_url, inpainting behavior

Model Notes

Nano Banana Pro

Strengths:

• lower-friction default for existing installs
• strong text rendering, including Cyrillic
• good at infographics, banners, and mixed text/image layouts
• supports --web-search in generate mode

Main params:

• --aspect-ratio
• --resolution
• --web-search (generate only)

OpenAI GPT Image 2

Strengths:

• stronger prompt adherence and fine-grained edits
• better photorealism and product-style renders
• native quality control
• uses the same FAL_KEY through fal, no separate OpenAI key
• current fal pricing is size/quality-dependent; a rough high-quality mental model is about $0.18 per image, but check the live model page before quoting an exact number

Main params:

• --image-size
• --quality

Compatibility layer:

• if --image-size is omitted, the scripts derive a valid OpenAI image_size from --aspect-ratio + --resolution
• this lets old prompts continue working after the provider switch in config

Workflow

Generate mode

1. Decide model from config
2. Clarify missing params only if needed:

• Nano Banana: aspect ratio, resolution
• GPT Image 2: image size and quality

1. Propose save path based on project structure
2. Run generate.sh
3. Parse result JSON, report URL and local files if downloaded

Edit mode

1. Get reference images:

• URL already available -> use directly
• local file -> upload.sh

1. Decide model from config
2. Clarify edit intent
3. Run edit.sh
4. Parse result JSON, report URL and local files if downloaded

Scripts

generate.sh

Nano Banana example:

sh scripts/generate.sh \
  --model "gemini" \
  --prompt "infographic about coffee brewing" \
  --aspect-ratio "9:16" \
  --resolution "1K" \
  --output-dir "./images" \
  --filename "coffee_infographic"

GPT Image 2 example:

sh scripts/generate.sh \
  --model "gpt" \
  --prompt "realistic product hero shot with sharp packaging text" \
  --image-size "landscape_4_3" \
  --quality "medium" \
  --output-dir "./images" \
  --filename "product_hero"

Compatibility example for GPT:

sh scripts/generate.sh \
  --model "openai" \
  --prompt "editorial portrait, window light, magazine cover layout" \
  --aspect-ratio "4:3" \
  --resolution "2K"

edit.sh

Nano Banana example:

sh scripts/edit.sh \
  --model "gemini" \
  --prompt "combine these into a collage" \
  --image-urls "https://example.com/img1.png,https://example.com/img2.png" \
  --aspect-ratio "16:9" \
  --output-dir "./images" \
  --filename "collage"

GPT Image 2 example:

sh scripts/edit.sh \
  --model "gpt" \
  --prompt "make this product shot look like a premium studio campaign" \
  --image-urls "https://example.com/source.png" \
  --mask-url "https://example.com/mask.png" \
  --image-size "auto" \
  --quality "medium" \
  --output-dir "./images" \
  --filename "studio_edit"

upload.sh

# Get hosted URL for local file
URL=$(sh scripts/upload.sh --file /path/to/image.png)
 
# Get base64 data URI for manual API work
URI=$(sh scripts/upload.sh --file /path/to/image.png --base64)

Cost Guidance

• Nano Banana Pro is the cheaper and safer default for quick iterations
• GPT Image 2 cost depends heavily on quality and image_size
• this skill defaults GPT to medium quality to reduce surprise spend

For current pricing, check fal's model pages in config/README.md.

Notes

• result URLs expire in roughly one hour — download locally if you need persistence
• uploaded files on fal storage are temporary
• edit.sh now polls the same /edit queue endpoints documented by fal