GPT Image Generate

If you want GPT's image generation - the one that's genuinely good at text-in-image and following long, messy prompts - this is the node: GPT Image Generate turns a prompt into an IMAGE tensor through OpenAI's API. But read the fine print: the API is called through the same 云雾 (Yunwu) relay that powers the rest of this pack, so there's no OpenAI account involved and no local model. You buy credits from the relay and it forwards your requests to the GPT image backend. The generation happens in someone's datacenter, not on your GPU.

How it works

Mechanically it's the most boring node in the pack, and that's a compliment. It takes your prompt and POSTs it to {api_url}/v1/images/generations with response_format: "b64_json" - a plain OpenAI-compatible images endpoint, same shape you'd hit with the official SDK. The response comes back base64-encoded, the node decodes it, converts it to a 0–1 float tensor, and stacks any results into a batch. Requests have a 120-second timeout, which matters because flagship image generation can genuinely take that long.

Inputs

• prompt - required, multiline. GPT image models reward detail; unlike Midjourney there are no flags to append.
• model - gpt-image-2-all (default), gpt-image-1.5, or gpt-image-1. The default tracks the pack's latest supported model, so leave it unless you know you need an older one.
• n - 1–10 images per call. Multiple results come back as a stacked batch in the single IMAGE output.
• quality - auto / high / medium / low.
• size - auto / 1024x1024 / 1536x1024 / 1024x1536. auto lets the model pick, which is usually the smart choice for text-heavy images.

One output: IMAGE, which wires straight into a preview or save node.

Installing

Same pack, same steps - ComfyUI Manager (search "ComfyUI-MidjourneyHub") or:

cd ComfyUI/custom_nodes
git clone https://github.com/jiaqianjing/ComfyUI-MidjourneyHub
# restart ComfyUI

The one difference: your key goes in config.ini under [GPT_IMAGE_API], not the Midjourney section. The repo ships the author's keys in that file, so replace them with your own before spending credits. The pack's requirements are light (aiohttp, openai, nest_asyncio); nothing model-sized to download.

Where people get burned

The relay's pricing groups are the classic trap. The README's own troubleshooting warns about "当前分组 ... 无可用渠道" errors - the relay only routes a given model through certain token groups, so a cheap group may simply have no channel for gpt-image-2-all or gpt-image-1.5, and you'll have to move to a pricier tier. Also set expectations: this is a paid, per-image service, so the "install it and generate forever" energy you'd bring to a local checkpoint doesn't apply - and the relay's README and console are in Chinese. It's a genuinely good model behind a slightly awkward middleman.