OpenAI API 连接 / OpenAI API Connector

The OpenAI API Connector is the workhorse of the RUI-Nodes pack's AI section: a node that sends text (and optionally up to six images) to any OpenAI-compatible chat-completions endpoint and returns the model's reply as a string. "Any compatible endpoint" is doing the heavy lifting here - the api_url and model fields are just strings, so besides OpenAI itself it works with DeepSeek, Moonshot, local Ollama servers, LM Studio, and every other API that speaks the OpenAI dialect. If you've got an LLM anywhere reachable by HTTP, this node can talk to it.

It earns its keep in ComfyUI workflows where you need a model to look at something: caption a generated image, compare two renders and pick the better one, summarize a prompt, or gate a step on a visual judgment. The image inputs are what make it interesting - this isn't just a text box, it's a vision-model client.

How it works

You give it a user_prompt (and optionally a system_prompt), and the node builds a chat request. The images are optional: plug image_1 through image_6 into whichever of the six inputs you have frames for. The detail here is thoughtful - each image is encoded separately and downscaled to a max long-edge of image_max_size (default 1024) before sending, which keeps token cost and payload size in check. And because each image is handled independently, they don't need to be the same resolution - a 512×512 and a 511×768 can be sent together, which standard batch-to-one-image nodes would choke on.

The inputs you'll actually set:

• api_url + protocol (https/http) - the endpoint. Default is api.openai.com/v1/chat/completions.
• api_key - your key. Saved in the workflow JSON, same as its sibling node - scrub before sharing.
• model - anything the endpoint accepts; default gpt-4o.
• user_prompt / system_prompt - what to ask.
• temperature (default 0.3), max_tokens (default 500), detail (low/high/auto) - standard API sampling knobs.
• proxy_url - for routing through a proxy, which the target audience for this pack will recognize as the difference between "works" and "connection refused."

Output is a single text string. Wire it into a text display, a save node, or back into a prompt.

Installing it

Part of RUI-Nodes ("Rui-Node🐶") - install once via ComfyUI Manager (search "RUI-Nodes") or:

cd ComfyUI/custom_nodes
git clone https://github.com/rui40000/RUI-Nodes
cd RUI-Nodes
pip install -r requirements.txt

Restart after. It needs requests and Pillow, both already in the pack's requirements.

Where people get burned

The usual cloud-API suspects: the key living in the workflow file, network egress needing a proxy in some regions, and cost on vision calls (images at high detail burn tokens fast - low or a smaller image_max_size helps). Also note the node deliberately clears proxy environment variables on load so a system-level proxy doesn't break requests - if your setup relies on an env proxy, use the proxy_url field instead of the environment. And one quiet behavior: if you leave user_prompt empty for pure text, some endpoints will just echo - give it a real prompt, the way you would any LLM.