ADIC OpenAIGPTImage1

This is a synchronous GPT Image 1 node that drops OpenAI's gpt-image-1 model straight into a graph, ComfyUI's comfy_api_nodes style: API_NODE = True, async under the hood, IMAGE tensor out. Wire a prompt to it, run, and a real image comes back - no local checkpoint, no VRAM cost, just a per-image API bill.

How it works is clean enough to be worth understanding. With only a prompt, it POSTs to /v1/images/generations (OpenAI-compatible JSON). The moment you also feed an image, it switches to /v1/images/edits as multipart/form-data - that's the GPT Image edit endpoint, and a batch of input images goes out as image[] fields. Add a mask on top and it becomes inpainting: per the tooltip, white areas get replaced. Inputs are downscaled to roughly 1.5MP before they're sent, so it won't choke on a 4K source.

Inputs that actually matter

• prompt - the only required input. Multiline.
• api_base - this is the one to watch. The default is https://api.gpt.ge, a third-party OpenAI-compatible proxy, not OpenAI's official endpoint. If you have your own OpenAI key and gateway, set this to https://api.openai.com/v1 or your relay. Most 401/authentication confusion with this node is people forgetting the base URL is a proxy.
• auth_token - Bearer token for the API. If your provider needs a Comfy Org API key instead, the node has a hidden key input for that path - either way, auth is "one bearer string, whichever the endpoint expects."
• quality - low / medium / high; the author's tooltip: "affects cost and generation time." Default low is a sensible place to start.
• background - opaque or transparent.
• size - auto, 1024x1024, 1024x1536, 1536x1024.
• n - how many images (1–8); more images, more money, and the output becomes a batch.
• image / mask - optional, turn text-to-image into edit/inpaint.
• seed - the author's own tooltip is blunt: "not implemented yet in backend." Leave it alone.

Output is a single IMAGE tensor (stacked, if n > 1), ready for a Preview or Save node.

Gotchas

The model name is hardcoded to gpt-image-1 - you can't swap it to a newer GPT Image generation without an update. And remember every generation costs real money on the API side; there's no "free tier" surprise here.

Installing

cd ComfyUI/custom_nodes
git clone https://github.com/jinchanz/ComfyUI-ADIC

Restart ComfyUI, or search "ComfyUI-ADIC" in ComfyUI Manager. No model files to download; the pack's README (Chinese, mostly about the image-translate nodes) only mentions requests, which ComfyUI already ships. If the node does nothing but error, check api_base and auth_token before anything else.