English | 中文
Generate images with the subscriptions you already have — no OPENAI_API_KEY.
A tiny zero-dependency Python CLI (and AI-agent skill): one file, stdlib only. It uses your ChatGPT subscription by default, falls back to Codex, and can use a Gemini subscription instead. Works on a free ChatGPT account too — the default backend just drives the normal ChatGPT web chat, where even free-tier users get image generation.
Formerly
chatgpt-imagegen. It was renamed once it grew non-ChatGPT backends. Thechatgpt-imagegencommand still works as an alias, and everyCHATGPT_IMAGEGEN_*environment variable is still honoured (the newIMAGE_USE_*name wins if both are set). Your saved styles are untouched.
image-use "a watercolor cat sitting on a windowsill" -o cat.png
# -> saved: cat.png (812,344 bytes) size=1024x1024 quality=medium
Needs Python 3.10+ and a ChatGPT subscription (free tier works).
For AI agents (recommended) — drops the skill into Claude Code, Codex, Cursor, etc.:
npx skills add leeguooooo/image-use -gThen just ask: "画一张 …" / "generate a hero banner for the README".
Standalone CLI — no pip, no virtualenv, no sudo. Keeps a git checkout under ~/.agents/use-family/image-use, links image-use into ~/.local/bin and the skill for Claude Code / Codex; re-run to update:
curl -fsSL https://raw.githubusercontent.com/leeguooooo/image-use/main/install.sh | shOr by hand:
git clone https://github.com/leeguooooo/image-use
sudo install image-use/image-use /usr/local/bin/image-use
sudo ln -sf image-use /usr/local/bin/chatgpt-imagegen # optional: keep the old command nameYou also need one backend — web (default, drives your logged-in Chrome, spends no Codex-usage) or codex (headless fallback). image-use doctor shows what's ready. → Backends & troubleshooting
Got a Gemini subscription too? Two more backends use it instead of OpenAI: --backend gemini (drives a logged-in gemini.google.com Chrome) and --backend agy (the Antigravity CLI, headless). They bill separate quotas from each other, so either can cover for the other. Neither is ever chosen automatically — ask by name. Pin the subscribed Chrome profile with --gemini-profile, since most profiles are signed in to some Google account. Note that Gemini text-to-image output carries a visible watermark in the bottom-right corner (image-to-image does not), and --size steers the aspect ratio there rather than the exact pixel count.
image-use upgrade # install the latest release and refresh the skill
image-use upgrade --check # only report: image-use 0.29.2 -> 0.30.0
image-use upgrade --json # the same as JSONupgrade (alias update) installs the newest GitHub release the same way this copy was installed — skills update for an npx skills add install, git pull --ff-only for a clone, a fresh copy of the script for a standalone file — then refreshes every other copy of the skill it finds (Claude Code plugin, clones and copies under ~/.agents/skills, ~/.claude/skills, ~/.codex/skills). Any other command checks for a newer release at most once a day and prints one line to stderr when there is one. IMAGE_USE_NO_UPDATE_CHECK=1 or the family-wide USE_NO_UPDATE_CHECK=1 turns the check off; it is also skipped when CI is set. Nothing is installed until you run upgrade.
On 0.23.1 or earlier? That self-update only looked for a global skills and gave up when it was missing, so it cannot deliver its own fix. Bootstrap once with (installs from before the rename are registered as chatgpt-imagegen):
npx -y skills update chatgpt-imagegenAfter that image-use upgrade works on its own.
image-use "moody mountain sunset" -o web/hero.png --size 1536x1024
image-use "make it a warm golden-hour photo, cinematic 35mm" -i photo.jpg # reference the subject
image-use "a different fictional person" --composition-ref photo.jpg # borrow the framing, not the face
image-use "a robot mascot" --style doodle # apply a gallery style (auto-pulled + saved)
image-use animate "a dog happily wagging its tail" --style-online snoopy --also-gif
OUT=$(image-use "icon" --quiet) # capture the path
image-use "product hero" --backend codex --image-model gpt-image-2.5-sunburst --quality xhighThe last line opts into the GPT Image 2.5 knobs — --image-model
(sunburst for precise editing, flare for fast high quality), --quality
(low→max), --background transparent, --compression, --action,
--partial-images. They are codex-only (the web/gemini surfaces have no
such controls), all optional, and are requests — the saved line prints the
model=/quality=/size= the backend actually used. Note: the codex
backend rewrites these server-side (observed gpt-image-2-codex / auto), so
they rarely survive there; what does matter on codex is --model, the driver
model that calls the tool (default gpt-5.6-luna, a fast/affordable Codex model
— a frontier coding model just burns the metered Codex bucket). Every run prints
the tokens= it cost.
All three below came straight out of the commands above — no retouching:
 |
 |
 |
"a watercolor cat sitting on a windowsill" |
"moody mountain sunset" --size 1536x1024 |
"a coffee shop logo, circular emblem" |
animate asks the image model for a strict 4×2 sprite sheet, crops eight equal
frames, rejects obvious subject drift, and writes a smooth ping-pong loop. It
defaults to animated WebP; pass --animation-format gif or --also-gif when
GIF compatibility matters. The source sprite PNG is always kept beside the
animation. Animation post-processing needs
ImageMagick (magick); WebP output additionally
needs libwebp (img2webp).
image-use doctor reports whether both are installed.
Full options: image-use --help. → Generate images · Styles
To organize web runs, --project accepts an exact name (found or created;
default imagegen) or an existing Project landing URL such as
https://chatgpt.com/g/g-p-<32-hex-id>/project. URL targets open directly by ID,
without listing or creating projects; localized display slugs, query strings,
and fragments are accepted and discarded when opening the canonical URL.
Malformed URL targets are rejected. --project "" uses a plain chat.
Project routing is best-effort by default: an unavailable project warns and
continues in a plain chat. Add --require-project (or IMAGE_USE_REQUIRE_PROJECT=1)
when that fallback would be unwanted. It requires a non-empty target and
--backend web or auto, verifies Project identity after opening and again in
a capture guard on the native Send click, and disables Codex fallback even when
the browser is unavailable. IMAGE_USE_PROJECT can supply either a name or URL.
Use --no-require-project to override the environment default for one run,
restoring best-effort routing and ordinary backend/fallback rules without
changing the exported variable.
Required-mode failures before Send retain the current draft for inspection. Inspect it in ChatGPT and manually clear the composer before the next run; ChatGPT can restore unsent drafts in new chats.
image-use "a watercolor cat" --project "Art" --require-project --keep-conversation
# To target an existing project exactly, set PROJECT_URL to its browser URL:
image-use "a watercolor cat" --project "$PROJECT_URL" --require-project --keep-conversationRouting and retention are independent: the conversation is deleted by default,
including with --require-project. Use --keep-conversation (or
IMAGE_USE_KEEP_CONVERSATION=1) to retain it; --keep-tab also retains it.
The ChatGPT browser backend pastes multiline prompts, checks the exact editor text, and waits for every reference upload to finish before clicking Send once. Incomplete uploads or altered text stop the run. If a send cannot be confirmed, it reports the uncertainty without sending again. Start with an empty composer; an existing draft is preserved. A run that stops before sending clears the text it pasted, except in required Project mode: it preserves the current draft for inspection because the route or draft ownership may have changed.
After sending, the web backend waits up to --timeout for a fresh image confirmed
in two consecutive page reads. Assistant text or a missing Stop control does not
prove image generation has finished. Detected rate-limit dialogs stop the run
immediately. When no image or streaming control is present, explicit English
reply prefixes such as "You've hit the image generation limit" or "I can't
generate that image." also fail immediately. Quoted mentions, vague "try again
later" text, and unrecognized replies keep waiting until the deadline; this is
not a complete quota/refusal detector for every wording or language. Timeout and
recognized reply-body errors include assistant text (up to 240 characters);
timeouts retain the last nonempty text through interrupted page reads.
On timeout, check the original conversation before retrying: the image may still
appear there, and auto does not fall back to Codex after submission. The error
names the conversation; once the image shows up, collect it without prompting
again:
image-use recover https://chatgpt.com/c/<id> -o out.pngrecover uses the chatgpt/images adapter from
chrome-use-sites (chrome-use site update
installs it) and the same single chatgpt.com tab and lock as a web run.
Browse and reuse art styles other people tuned — a public gallery at drawstyle.leeguoo.com. No script update needed:
image-use "a fox barista" --style-online doodle # generate with a gallery style, nothing saved
image-use style search "watercolor mascot" # search the gallery
image-use style publish mystyle --category cute --from-last # share yours (one-time login)
image-use upload out.png --style doodle # share a result to the style's player gallery (no login, on request)A style can pin a character, not just a look. Style assets carry reference images, so the same character comes back in a brand-new scene:
image-use style add pip --kind character --ref pip-ref.png
image-use "a fox barista" --style pip |
 |
| the pinned reference | generated from "a fox barista" |
Gallery packages can be characters too — xiaohei is one.
→ Using gallery styles · Submitting a style
- 📖 Full documentation — install, generating, styles, backends, the platform.
- 🎨 Style gallery — browse and contribute community art styles.
- 📝 Deep dive (blog) — the design and principles behind it.
- ⚙️ How it works · HTTP API wrapper
- 🚀 Releasing:
scripts/release.sh <version> "<whatsnew>"(--dry-runto preview) bumps the version, tests, pushesmain, waits for the CI-created Release, then syncs the plugin marketplace.
MIT — see LICENSE.
This tool calls ChatGPT's internal backend-api/codex endpoint — the same one the official Codex CLI uses. It is not a documented public API; OpenAI could change or restrict it at any time. Use at your own risk and within the OpenAI Terms of Use — in particular, do not use your ChatGPT subscription to power a public-facing image generation service.
Keywords
ChatGPT subscription image generation, free ChatGPT account image generation, use ChatGPT Plus for image API, gpt-image-2.5 without OPENAI_API_KEY, gpt-image-2.5 ChatGPT subscription, gpt-image-2 ChatGPT subscription, image_generation tool Responses API, ChatGPT image CLI, Codex CLI image_gen as standalone tool, DALL-E via ChatGPT Plus, OAuth-backed OpenAI image generation, no-API-key image generation, AI agent image generation skill, Claude Code image skill, OpenAI image generation without billing.
中文： 用 ChatGPT 订阅生成图片、免费 ChatGPT 账号生图、ChatGPT Plus 生图工具、不用 API key 生图、gpt-image-2 用订阅、ChatGPT 订阅生图 CLI、Codex CLI 生图能力独立工具、给 AI agent 用的生图 skill、本地生图脚本、零依赖 Python 生图工具。
Built by 郭立 (Guo Li / leeguoo) — leeguoo.com · GitHub · X · more tools in the *-use family.
Use --backend codex --codex-provider current (or set
IMAGE_USE_CODEX_PROVIDER=current) to select the top-level model_provider
from $CODEX_HOME/config.toml (default ~/.codex/config.toml). A provider name
such as --codex-provider myrelay selects [model_providers.myrelay] directly.
Provider mode requires Python 3.11+; leaving the option/environment variable
empty preserves the existing ChatGPT OAuth backend on Python 3.10+.
model = "gpt-6-astra"
model_provider = "myrelay"
[model_providers.myrelay]
base_url = "https://relay.example"
wire_api = "responses"
requires_openai_auth = false
auth = { command = "/usr/local/bin/get-token", args = ["--name", "relay"], timeout_ms = 15000, refresh_interval_ms = 3600000 }Recommended: use auth.command, which also works in desktop Codex and keeps the
returned token out of config files and environment variables. The command and
args run directly without a shell; trimmed stdout becomes the Bearer token.
auth must be a table with a nonempty string command. Optional args must be
a string list; timeout_ms must be a positive integer (default 15000); optional
string cwd sets the working directory. This one-shot CLI ignores
refresh_interval_ms and caches command success or failure once per provider
per process. Priority is env_key → auth.command → experimental_bearer_token;
failure of the selected source does not fall back. Alternatively, set
env_key = "MYRELAY_API_KEY" and inject the token securely into the process
environment; avoid plaintext experimental_bearer_token in config. The endpoint is
base_url.rstrip("/") + "/responses"; the relay must support the Responses
image_generation tool. Provider mode uses no ChatGPT OAuth or account ID.
The top-level model supplies the driver default; --model (or
IMAGE_USE_MODEL) overrides it. Static http_headers and env_http_headers
(header name → environment variable name) are supported; unset header
variables are skipped. image-use doctor --codex-provider current checks
configuration/token readiness without displaying the token.