openai-imagegen

name	openai-imagegen
description	Generate and edit images using OpenAI's gpt-image-2 model — the highest-quality image generation model available (Arena

OpenAI Image Generation (gpt-image-2)

Generate and edit images using OpenAI's gpt-image-2 API — the highest-quality image model available as of April 2026.

Before You Start

1. Check if OPENAI_API_KEY is set in the environment.

2. If OPENAI_API_KEY is NOT available → use ChatGPT Prompt Mode. Instead of running scripts, output a ready-to-paste prompt the user can copy into ChatGPT. Follow these rules:

   • Output the full image generation prompt in a code block labeled "ChatGPT Prompt"
   • Apply all the prompting best practices from the Prompting Guide below — structure it as background → subject → details → constraints
   • Include recommended settings as a note: quality, size, aspect ratio
   • If the task involves editing, tell the user which image(s) to attach and what instruction to give
   • If the task involves multiple images (e.g., a series), give each prompt separately numbered

3. If OPENAI_API_KEY IS available → use the scripts. Use the venv Python for all script invocations:

SCRIPTS_DIR="${CLAUDE_PLUGIN_ROOT}/skills/openai-imagegen/scripts"
PYTHON="${SCRIPTS_DIR}/.venv/bin/python3"

If the venv doesn't exist yet, create it:

python3.12 -m venv "${SCRIPTS_DIR}/.venv"
"${SCRIPTS_DIR}/.venv/bin/pip" install openai Pillow

Scripts

Use these scripts for all image generation tasks. Always invoke with $PYTHON:

generate_image.py — Create images from text

Use when: User wants to create a new image from a text description

$PYTHON "${SCRIPTS_DIR}/generate_image.py" "prompt" output.png [--quality QUALITY] [--size SIZE] [--format FORMAT] [--preset PRESET] [--n COUNT]

edit_image.py — Modify existing images

Use when: User wants to edit an image, use reference images, or do inpainting with a mask

# Edit with instruction
$PYTHON "${SCRIPTS_DIR}/edit_image.py" "instruction" output.png --image input.png [--quality QUALITY] [--size SIZE]

# Edit with mask (inpainting)
$PYTHON "${SCRIPTS_DIR}/edit_image.py" "instruction" output.png --image input.png --mask mask.png

# Compose from multiple reference images
$PYTHON "${SCRIPTS_DIR}/edit_image.py" "instruction" output.png --image ref1.png --image ref2.png

Script Options

Option	Values	Default
--quality	low (drafts, fast), medium (default, good balance), high (print-ready — only use when explicitly requested), auto	medium
--size	Any WxH where both are multiples of 16, max edge 3840px, aspect ≤3:1. Or auto	auto
--format	png, jpeg, webp	png
--compression	0-100 (for jpeg/webp only)	None
--n	1-8 (number of images)	1
--preset	draft, social, print, wide, portrait	None

Presets

Preset	Quality	Size	Format	Best for
draft	low	1024x1024	jpeg	Quick iterations, brainstorming
social	medium	1080x1080	png	Instagram, social posts
print	medium	2048x2048	png	Print materials, high-res assets
wide	medium	1920x1080	png	Presentations, headers, banners
portrait	medium	1080x1920	png	Stories, mobile, posters

Output

Images are saved to the specified output path. For Slack sharing, save to ~/work/ or /tmp/ and use the attach: prefix in your response:

Here's your image attach:/tmp/generated-image.png

When generating multiple images (--n > 1), files are saved as output_1.png, output_2.png, etc.

Prompting Guide

Forget old limitations

gpt-image-2 is a generational leap. It has native reasoning — it thinks about your prompt before drawing. Throw away every instinct from DALL-E, Midjourney, or Stable Diffusion about what AI image generation "can't do." This model:

• Renders text perfectly — full paragraphs, menus, labels, code snippets, multilingual
• Places elements exactly where you tell it to
• Maintains character consistency across multi-panel outputs
• Understands complex spatial layouts, infographics, data visualizations
• Handles dense compositions with dozens of elements
• Searches the web for reference before generating (in thinking mode)

Be as specific and ambitious as you want. Describe exactly what you see in your head. The more precise and detailed your vision, the better the output. Don't simplify for the model — it can handle it.

Prompt structure

Order your prompt for clarity:

1. Format and dimensions — What is this? (poster, banner, infographic, comic, UI mockup)
2. Layout — Exact placement of elements (top-left, centered, grid arrangement)
3. Content — Every piece of text, every visual element, every icon
4. Style — Colors (use hex codes), typography (font styles, weights, sizes), mood, aesthetic references
5. Constraints — What to exclude, background treatment

Text in images

Just write the exact text you want. The model renders it accurately. For best results:

• Put literal text in "quotes" in the prompt
• Specify font characteristics: "bold sans-serif", "thin monospace", "elegant serif italic"
• Specify placement: "centered at the top", "bottom-right corner", "along the left margin"
• For complex layouts with multiple text blocks, describe each one with its position, size, and style

Be bold

• Want a full restaurant menu with 30 items, prices, and descriptions? Ask for it.
• Want an infographic with real data labels, axis markers, and a legend? Describe it.
• Want a comic strip where the character looks identical in every panel? Just say so.
• Want a magazine cover with a headline, subtitle, barcode, and pull quotes? Go for it.

Iteration

Default is medium quality — good enough for most uses at ~$0.05/image. Only bump to high when the user explicitly asks for print-ready or highest quality output. Use low for rapid iteration and brainstorming.

When to Use This vs Gemini

Scenario	Use
Highest quality, client-facing assets	OpenAI (this skill)
Accurate text in images	OpenAI
Multi-panel with character consistency	OpenAI
Quick/cheap generation, exploration	Gemini
Speed-critical (1-3s generation)	Gemini
4K output	Gemini (pro model)
Budget-sensitive bulk generation	Gemini

Cost Reference

At 1024x1024:

• Low: ~$0.006/image
• Medium: ~$0.053/image
• High: ~$0.211/image

Default quality is medium (~$0.05/image). Only use high when the user explicitly requests it.