Zhenzhen_gpt_image_1

If you've ever needed an image with legible text in it - a poster, a sign, a menu, a screenshot-looking mock - you know GPT Image is in a different league from most open models, where the text comes out as squiggles. Comfly_gpt_image_1 puts that capability in a ComfyUI node: text-to-image against gpt-image-1 (or the newer gpt-image-1.5), called through the Zhenzhen relay. It's a plain single-prompt node, no reference-image editing here (that's the _edit sibling's job).

One thing to know about this node specifically: the relay's default token group for gpt-image-1 is a reverse-engineered cheap lane - works, but slow and occasionally flaky. The official groups (the ssvip one is called out as the best value) are faster and stable but cost more. The README says it plainly: cheap lane = "按次收费，缺点就是不稳定，速度慢." Pick your lane on the website under your token's groups.

The inputs that matter

• prompt - what you're making. For GPT Image, be specific about the text you want rendered and its placement.
• model - gpt-image-1 or gpt-image-1.5. 1.5 is the newer, higher-fidelity release; 1 is the battle-tested default.
• n - how many images (1–10). Each one bills.
• quality - auto, high, medium, low. low is genuinely useful for fast drafts; high for the deliverable.
• size - auto, or explicit 1024x1024, 1536x1024, 1024x1536.
• background - auto, transparent, or opaque. Transparent backgrounds are one of GPT Image's quiet superpowers for asset work.
• output_format - png, jpeg, or webp.
• moderation - auto or low; the latter relaxes the content filter slightly if the relay passes it through.

Outputs: generated_image (IMAGE tensor) and response (JSON).

Install and key

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/Comfyui-zhenzhen

Restart, or ComfyUI Manager → "Comfyui-zhenzhen". API key from ai.t8star.org into api_key. No models to download.

Caveats that bite

The cheap default group is the classic "it's cheap because it's reverse-engineered" trade - great for testing, frustrating for production when it stalls. If you're building something serious on this node, move your token to an official group. OpenAI's moderation is the filter that actually applies, so don't plan around bypassing it. And the per-call meter runs on every image - a 10-image batch is ten charges. For a node that can render a crisp billboard sign on the first try, that's a fair price; just don't treat it as free.