Image Generation

Image Generation

Generate images using Google's Gemini model (default) or fal.ai FLUX.2 klein 4B (cheap mode).

Quick Start

MANDATORY: Always use -d (destination) flag to specify output path. This avoids file management issues with poster_0.jpg in the scripts folder.

cd ~/.claude/skills/image-generation/scripts

# ALWAYS specify destination with -d flag
npx ts-node generate_poster.ts -d /tmp/my-image.jpg "A futuristic city at sunset"

# Cheap mode: fal.ai FLUX.2 klein 4B (fast, lower cost)
npx ts-node generate_poster.ts --cheap -d /tmp/city.jpg "A futuristic city at sunset"

# With aspect ratio
npx ts-node generate_poster.ts -d /tmp/landscape.jpg --aspect 3:2 "A wide landscape poster"
npx ts-node generate_poster.ts --cheap -d /tmp/story.jpg -a 9:16 "A vertical story format"

# With reference assets (image editing)
npx ts-node generate_poster.ts -d /tmp/banner.jpg --assets "/path/to/avatar.jpg" "Create banner with character"
npx ts-node generate_poster.ts --cheap -d /tmp/photo.jpg --assets "/path/to/image.jpg" "Turn this into a realistic photo"

Why -d is mandatory:

• Output goes to predictable location (e.g., /tmp/ for temp images)
• Avoids hunting for poster_0.jpg in scripts folder
• Enables immediate use without file renaming

Providers

Cheap Mode (--cheap)

Uses fal.ai FLUX.2 klein 4B - a distilled FLUX model optimized for speed and cost.

Why use cheap mode?

• Cost: ~$0.003 per image (vs Gemini's higher cost)
• Speed: Fast 4-step inference (~2-4 seconds)
• Quality: Good enough for drafts, social media, iterations
• Batch friendly: Generate many variations quickly

Endpoints:

• Text-to-image: fal-ai/flux-2/klein/4b
• Image editing (with --assets): fal-ai/flux-2/klein/4b/edit

Best for:

• Quick iterations and previews
• Social media content
• Concept exploration
• Batch generation (10+ images)

Use Gemini instead for:

• Final production assets
• Complex compositions
• Text-heavy images (especially Hebrew)

Quality

Control image resolution with --quality or -q:

npx ts-node generate_poster.ts -q 2K "High quality poster"

Aspect Ratio

IMPORTANT: Always use the default 3:2 aspect ratio unless the user explicitly requests a different format (like "vertical", "story", "square", etc.). Do NOT change the aspect ratio on your own.

Control image dimensions with --aspect or -a:

npx ts-node generate_poster.ts --aspect 3:2 "Your prompt"
npx ts-node generate_poster.ts -a 16:9 "Your prompt"

Adding Assets (Reference Images)

Use --assets with full paths to include reference images:

# Single asset
npx ts-node generate_poster.ts --assets "/full/path/to/image.jpg" "Your prompt"

# Multiple assets (comma-separated)
npx ts-node generate_poster.ts --assets "/path/a.jpg,/path/b.png" "Use both images"

# With cheap mode - uses fal.ai edit endpoint
npx ts-node generate_poster.ts --cheap --assets "/path/to/image.jpg" "Turn this into a painting"

Supported formats: .jpg, .jpeg, .png, .webp, .gif

IMPORTANT: Assets are NOT automatically included. You must explicitly pass them via --assets.

Destination (MANDATORY)

ALWAYS use --destination or -d to specify output path:

# REQUIRED: Always specify destination
npx ts-node generate_poster.ts -d /tmp/poster.jpg "My poster"

# If file exists, auto-adds suffix: poster_1.jpg, poster_2.jpg, etc.

Features:

• Auto-creates parent directories if needed
• Collision avoidance: if file exists, adds _1, _2, etc.
• Use /tmp/ for temporary images that will be uploaded elsewhere

Save to Gallery

Save good results for future style reference:

npx ts-node generate_poster.ts --save-to-gallery "my-style" "prompt"

Creates assets/gallery/my-style.jpg + .meta.json with prompt info.

API Configuration

Create scripts/.env:

GEMINI_API_KEY=your_gemini_api_key
FAL_KEY=your_fal_api_key

Hebrew/RTL Content

IMPORTANT: When the image contains Hebrew text, you MUST add the following sentence to the prompt:

"CRITICAL: Layout is RTL (right-to-left). All text in Hebrew. Visual flow, reading order, and panel sequence go from RIGHT to LEFT."

Copy this exact sentence and paste it at the BEGINNING of your prompt. Without it, the image will render left-to-right which is wrong for Hebrew content.

WOW Mode

When user asks for "wow mode" or wants maximum visual impact, add these epic elements to the prompt:

"EPIC CINEMATIC WOW: Amazed expression, mind-blown reaction, intense VFX - shattered glass particles, explosive energy bursts, volumetric light rays, dramatic lens flares, particle explosions, motion blur streaks, holographic glitches, electric sparks, cinematic color grading, dramatic rim lighting, depth of field bokeh, anamorphic lens effects. Maximum visual spectacle."

This pushes the visual to the edge with:

• Shattered/particle effects
• Explosion/energy bursts
• Lens flares and volumetric lighting
• Motion blur and dynamic effects
• Electric/holographic glitches
• Dramatic cinematic lighting

Output

• Files saved to path specified by -d flag (MANDATORY)
• Aspect ratio: Configurable via --aspect (default: 3:2)
• Quality: 1K (1024px on longest edge)
• Send via WhatsApp after generation (if whatsapp skill exists)