STOP. READ THIS ENTIRE SKILL.MD BEFORE CALLING ANY ENDPOINT.

create-image

Generate images using FREE AI image generation backends.

Features

• Ollama (local) - Z-Image Turbo or FLUX2-Klein via Ollama (FREE, no internet)
• Gemini 2.5 Flash Image - AI-generated images via Google gemini-2.5-flash-image (FREE with API key)
• FLUX.1-schnell - AI-generated images via HuggingFace (FREE remote)
• Mermaid diagrams - Flowcharts and architecture diagrams (FREE)
• Placeholder images - Random grayscale from picsum.photos (FREE)
• Solid color - Gray box with text label (always works)
• Size control - Specify dimensions for PDF embedding (auto-resized)

Quick Start

cd .pi/skills/create-image

# Generate an AI image (uses Gemini or FLUX)
uv run --script generate.py "hardware verification flowchart for microprocessor" \
  --output test_figure.png \
  --size 400x600

# Generate with specific backend
uv run --script generate.py "network security architecture" \
  --output security_arch.png \
  --size 800x600 \
  --backend flux

# Use placeholder fallback
uv run --script generate.py "placeholder" \
  --output placeholder.png \
  --size 400x300 \
  --backend placeholder

Commands

generate - Create an image

uv run --script generate.py "<prompt>" [options]

Arguments:

Options:

Backends

Setup

Option 1: Ollama (macOS only - MLX framework)

Note: Ollama image generation currently only works on macOS (Apple Silicon). Linux/NVIDIA support is "coming soon" per Ollama docs.

# macOS only
ollama pull x/z-image-turbo
# or
ollama pull x/flux2-klein

Option 2: Google Gemini 2.5 Flash Image (Nano Banana) (FREE API)

Get a FREE API key from aistudio.google.com:

export GEMINI_API_KEY="your_api_key_here"
# or
export GOOGLE_API_KEY="your_api_key_here"

Note: This uses the gemini-2.5-flash-image model (aka "nano-banana") via the REST API. No special SDK installation required (uses requests). Image generation counts against your daily Pro quota (~1000 images/day). Either GEMINIAPIKEY or GOOGLEAPIKEY will work.

Option 3: HuggingFace Token (FREE Remote)

Get a FREE HuggingFace token from huggingface.co/settings/tokens:

export HF_TOKEN="hf_your_token_here"

Option 4: Mermaid (for Diagrams)

npm install -g @mermaid-js/mermaid-cli

Example Prompts

Security Documents

"APT attack kill chain diagram with reconnaissance, weaponization, delivery, exploitation phases"
"network intrusion detection system architecture"
"malware analysis workflow flowchart"

Engineering Documents

"hardware verification flow for microprocessor with RTL, synthesis, and timing analysis"
"FPGA design pipeline from HDL to bitstream"
"embedded systems boot sequence diagram"

Scientific Documents

"machine learning pipeline with data preprocessing, training, and inference stages"
"experimental methodology flowchart"
"system architecture diagram with numbered components"

Cached Images (Reuse Before Generating)

Pre-generated images are available in cached_images/ - use these first to avoid unnecessary API calls:

# Copy cached image instead of generating
cp cached_images/flowchart.png /path/to/output.png

Integration with PDF Generation

After generating images, embed them in PDFs:

import fitz  # PyMuPDF

doc = fitz.open()
page = doc.new_page()

# Insert generated image
img_rect = fitz.Rect(50, 200, 450, 500)  # x0, y0, x1, y1
page.insert_image(img_rect, filename="test_figure.png")

doc.save("fixture_with_figure.pdf")

Common Mistakes

WRONG: Using AI generation for UI mockups (garbled text, stretched layouts)

uv run --script generate.py "dashboard with sidebar and data table" --output mockup.png
# Diffusion models produce unreadable text in UI layouts

RIGHT: Use AI generation only for icons, logos, artwork. Use HTML/CSS for UI mockups

# For UI: write HTML/CSS, render to PNG via browser screenshot
# For artwork: AI generation is fine
uv run --script generate.py "abstract nebula background" --output nebula.png

WRONG: Generating images when cached versions exist

uv run --script generate.py "technical workflow diagram" --output flow.png
# Wastes API calls when cached_images/flowchart.png already exists

RIGHT: Check cached_images/ first

cp cached_images/flowchart.png /path/to/output.png

WRONG: Not specifying size for PDF embedding

uv run --script generate.py "figure" --output fig.png  # 512x512 default, may not fit

RIGHT: Specify dimensions matching the target layout

uv run --script generate.py "figure" --output fig.png --size 400x600

Dependencies

dependencies = [
    "huggingface_hub>=0.26.0",
    "httpx",
    "typer",
    "pillow",
]