GPT Image 2

Use the bundled Node.js CLI for lightweight image generation and editing. It has no npm dependencies and supports OpenAI-compatible image endpoints.

Quick Start

Run commands from this skill folder:

node scripts/gpt-image2.mjs --help
node scripts/imgbb-upload.mjs --help

Configure an API channel:

node scripts/gpt-image2.mjs config set openai --api-key "$OPENAI_API_KEY" --base-url "https://ai-api.qzsyzn.com/v1" --model "gpt-image-2" --edit-transport generations
node scripts/gpt-image2.mjs config use openai

Use --edit-transport multipart for the official OpenAI-style POST /images/edits multipart flow (default when omitted).

For unified JSON edit (generations), do not rely on huge base64 payloads. Prefer HTTPS URLs:

Default: local reference files are uploaded to ImgBB automatically (built-in API key, stored encrypted in scripts/imgbb-upload.mjs). No need to set IMGBB_API_KEY unless you want your own key—then use IMGBB_API_KEY or --imgbb-key (see ImgBB API).
Or upload first, then edit with URLs only:

node scripts/imgbb-upload.mjs ./test/image.png
node scripts/gpt-image2.mjs edit --edit-transport generations --image "https://i.ibb.co/..." --prompt "Your edit instruction" --output outputs/edited.png

Global install after npm link: imgbb-upload ./photo.png prints one URL per line.

Use --embed-local-base64 only when you intentionally want data URLs (e.g. offline debugging).

Try a prompt template:

node scripts/gpt-image2.mjs templates list
node scripts/gpt-image2.mjs try encyclopedia-card --var topic="coffee brewing" --output outputs/coffee.png

Query supported resolutions:

node scripts/gpt-image2.mjs sizes list
node scripts/gpt-image2.mjs sizes list --orientation portrait

Generate with a direct prompt:

node scripts/gpt-image2.mjs generate --prompt "A clean product hero image of a ceramic desk lamp" --size 1024x1024 --output outputs/lamp.png

Edit an existing image (generations transport uploads local files via ImgBB by default):

node scripts/gpt-image2.mjs edit --image input.png --prompt "Keep the object unchanged, replace the background with a warm studio setup" --output outputs/edited.png --edit-transport generations

Optional: export IMGBB_API_KEY="..." or --imgbb-key ... to use your own ImgBB key instead of the built-in default.

Print a curl command instead of calling the API:

node scripts/gpt-image2.mjs generate --template city-poster --var city="Chengdu" --curl

Workflow

Use templates list, templates search <query>, and templates show <id> before drafting prompts from scratch.
Prefer try <template-id> --var key=value for quick experimentation.
Use config set <channel> for provider-specific API keys, base URLs, and model names. The active channel is used by default.
Keep OpenAI-compatible API keys in config or environment variables, not in checked-in files. ImgBB uses a built-in default unless you override with IMGBB_API_KEY or --imgbb-key.
Run sizes list before choosing --size, especially when switching between square, portrait, and landscape output.
Use --dry-run to inspect the final payload and --curl when the user wants shell-native invocation (with ImgBB enabled, dry-run still contacts ImgBB so URLs in the JSON match a live run).
If an API call fails once, report the HTTP status and response body. Do not blindly retry without changing prompt, size, channel, or timeout.

Configuration

Credential precedence:

CLI flags: --api-key, --base-url, --model, --edit-transport, --imgbb-key
Active channel in the config file
Environment variables: GPT_IMAGE_API_KEY, OPENAI_API_KEY, API_KEY; GPT_IMAGE_BASE_URL, OPENAI_BASE_URL, BASE_URL; GPT_IMAGE2_EDIT_TRANSPORT (multipart or generations); IMGBB_API_KEY (optional; overrides the built-in ImgBB key when uploading locals for generations edit)

The config file defaults to ~/.gpt-image2/config.json. Override it with GPT_IMAGE2_CONFIG.

Useful commands:

node scripts/gpt-image2.mjs config list
node scripts/gpt-image2.mjs config show
node scripts/gpt-image2.mjs config use <channel>
node scripts/gpt-image2.mjs config set <channel> --api-key "..." --base-url "..." --model "..." [--edit-transport multipart|generations]

Prompt Templates

Templates live in references/prompts.json. Keep templates reusable:

Use {{variable}} placeholders for user-specific details.
Avoid names or likenesses of real private people unless the user explicitly provides appropriate rights and context.
Prefer complete visual specifications: subject, scene, composition, lighting, style, details, output use.
Keep new templates as compact patterns rather than copying long community prompts verbatim.

API Notes

The CLI calls:

POST /images/generations for text-to-image (JSON).
Edit mode (default): POST /images/edits with multipart form data (--edit-transport multipart).
Unified gateways: same JSON POST /images/generations with an image array of strings. Prefer HTTPS URLs: by default, local files are uploaded via ImgBB (built-in key); only URLs are sent in JSON. Use --embed-local-base64 to force data URLs instead. Optional IMGBB_API_KEY / --imgbb-key overrides the default ImgBB key. Use scripts/imgbb-upload.mjs for a standalone upload step. Up to 5 reference images; --mask is not supported in this mode.

It saves b64_json image data when present and can also download a returned url for OpenAI-compatible gateways that use URLs.