URL图像编辑丨API

The OpenAI images/edit endpoint - the one that takes a picture, an instruction, and optionally a mask, and returns an edited image - is the shape behind a lot of hosted inpainting and style work. CustomURLImageEditAPI is the QING node that exposes it against any OpenAI-compatible service: an official OpenAI key, a Chinese gateway that mirrors the format, or a self-hosted proxy. It sends your image(s) and prompt as a multipart edit request, and if you wire a mask in, that becomes an inpainting-style edit where only the masked region changes.

It's part of ComfyUI-QING's API category. Unlike the language/vision chat nodes, this one returns images, not text: image comes back as an IMAGE tensor ready to feed a preview or save node.

How it works. The node builds a multipart request to your base_url (root or /v1 accepted) using the model and prompt you supply. Details worth knowing from the source: when you connect multiple images (image2, image3), they're submitted as image[] fields; the optional mask goes out as a mask field - that's the inpainting hook. size (default 1024x1024) sets the output dimensions and is omitted from the request if you leave it blank. n controls how many images come back (1–4) and returns them as an IMAGE batch. response_format is the fork: b64_json (default) expects the image inline as base64, url expects the service to return a URL the node fetches - pick whichever your endpoint actually speaks.

The inputs that matter.

• prompt - the edit instruction.
• image - required; the primary image to edit.
• base_url and model - your endpoint and its model ID.
• size, n, response_format - the output controls above.
• api_key_override - set it per-node, or rely on CUSTOM_URL_API_KEY / QING settings.
• Optional: image2/image3, mask, custom_headers_json, extra_body_json, timeout_sec (180).

Outputs: image (IMAGE), raw_response, status, debug_info.

How to install. ComfyUI-QING is a one-pack install - search "ComfyUI-QING" in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/GAO-SHIQING/ComfyUI-QING
cd ComfyUI-QING
python install_dependencies.py

The openai dependency is in the pack's requirements. Restart after installing. The README's clone URL is typo'd (GAOSHI-QING, missing the H) - use the URL above or Manager.

Troubleshooting. The response_format mismatch is the first thing to check when you get an empty image: if the service returns URLs but the node is set to b64_json, nothing decodes. Read status/debug_info - that's what they're for. Second: mask support is endpoint-dependent; if your service doesn't accept a mask field, it'll reject the request, and you're doing straight edits instead of inpainting. Third, edit endpoints are picky about size - some models only accept their own supported sizes, so a custom size that the server rejects will 400; blank it to let the server default. And the usual API-node reminders: your images leave the machine to the endpoint, the service's content policy applies, and per-call cost adds up faster than you think if you're iterating edits in a loop.