arkennemasis Codex Image Gen (ChatGPT login)

The name is the pitch: ArkCodexImageGen generates or edits images with OpenAI's gpt-image-2 inside ComfyUI, using your existing ChatGPT/Codex login instead of an API key. If you already pay for ChatGPT, you get gpt-image-2 generations billed against your plan rather than per call. No API key field exists. No OAuth flow exists inside ComfyUI - the node reads the login the Codex CLI already wrote on this machine.

That "no API key" claim is the headline, but the honest caveat sits right next to it: it requires a paid ChatGPT plan (the free tier can't call the hosted image tool), and you have to install the Codex CLI and run codex login once in your own terminal:

codex login

That writes ~/.codex/auth.json, which the node reads directly. codex logout and the node stops working; log in as a different account and it follows. It's one login, reused - the same OAuth credentials, not a copy. The pack also ships ArkCodexLoginStatus so you can confirm the account and plan before spending anything.

The inputs that matter

• prompt - required, your text-to-image prompt or the edit instruction.
• image_1 through image_4 - optional reference/edit images (can be batches). This is the edit path: give it an image and a "change the background to X" prompt.
• aspect_ratio - stated in the prompt, because this backend ignores the size field. The tooltip is blunt about the tradeoff: wide/tall targets land close but not exact (16:9 came back as 7:4). If the ratio must be precise, use the pack's Replicate node instead - that's the honest recommendation, and it's why both paths exist.
• quality and background - defaults, low/medium/high, and opaque/transparent/auto.
• codex_home - point at a specific account's auth folder for multi-account setups. Blank = $CODEX_HOME, else ~/.codex.
• allow_refresh - default on: refresh an expired token and save it back atomically. Off = fail instead of writing to auth.json.
• run_mode / max_concurrent - one at a time (default) serialises every arkennemasis API node via an asyncio lock; all at once with max_concurrent (default 2) runs calls concurrently. Under $5 credit Replicate drops to ~6 req/min, so parallel calls without a cap reliably 429 - that's why the lock exists.
• on_failure - what to do when the backend definitively returns no image after every retry (usually a moderation refusal). fail the run is right for a single image; inside a 24-scene loop one refusal on the 23rd image threw away 22 finished images and 36 minutes of work, which is why use the reference image exists - it substitutes the input image so the run completes, and says so loudly in the log.
• settings - wire the shared Image Gen Settings node here to drive several generators from one place.

Outputs: image (the IMAGE tensor) and account (the signed-in email that produced it - the way to tell which login made which image in a multi-account graph).

Installing and expectations

It's part of comfyui-arkennemasis:

cd ComfyUI/custom_nodes
git clone https://github.com/Hishamahmer/comfyui-arkennemasis
pip install replicate httpx

restart, and it's under arkennemasis/Image Gen. It needs the httpx dependency to load, and the Codex CLI + a paid plan to run.

Where people get burned: expecting pixel-exact aspect ratios (use Replicate for that), running a batch without checking the login first, and forgetting that availability is account-dependent - not every plan or region can call the hosted image tool, and the node says so plainly rather than dumping an HTTP error. Also note the node distinguishes a finished image from a partial_images preview, so a stream cut mid-render can never be saved as if it were the final result. That's the kind of defensive detail that makes a paid-API node trustworthy.