URL视觉丨API

Where the pack's named vision nodes are locked to one provider, CustomURLVisionAPI is the open door: give it an OpenAI-compatible base_url, a model ID, and it becomes a vision node for whatever VLM that endpoint serves - a local vLLM running Qwen-VL, an OpenRouter vision model, a hosted gateway. If you've already got one working OpenAI-compatible vision endpoint, this is the node that makes the QING pack feel worth installing by itself.

It's part of ComfyUI-QING's API category, and it shares the multi-turn history machinery of the pack's other chat nodes. The distinguishing feature over the plain language node is, obviously, images: it accepts up to three (via image2 and image3) and sends them along with your question for multi-image understanding.

How it works. Standard OpenAI chat-completions client pointed at your base_url (root or /v1 both accepted), with your image attached and the question in text_input. The image_detail enum (auto/low/high) maps to OpenAI's image_url.detail parameter - that's the "how much of the image do we actually send at high resolution" control, and low is a meaningful speed/cost win for simple checks. Keys resolve in the same priority order as the pack's other custom-URL nodes: api_key_override, then CUSTOM_URL_API_KEY, then the QING settings panel. History defaults to 6 rounds - enough to refine a description of the same image, short enough that captions don't go stale.

The inputs that matter.

• image - required; what the model sees.
• text_input - your question (default "describe this image").
• base_url and model - the endpoint and model ID, exactly as your server names them.
• max_tokens (2048) and history (6).
• Optional: image2/image3, image_detail, temperature, top_p, repetition_penalty, timeout_sec, clear_history, plus custom_headers_json and extra_body_json for nonstandard auth or params.

Outputs: generated_text, conversation_info, total_tokens, raw_response, status - the same five-port bundle as the language node, with raw_response/status as your debugging window.

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 library is already 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. Same endpoint-debugging playbook as CustomURLLanguageAPI: wrong base_url, wrong model ID, or a key not where the node looks - and status/raw_response will tell you which. Vision adds two quirks: some cheap/old OpenAI-compatible endpoints don't support image_detail, so set it to auto if you get parameter errors; and if your endpoint rejects repetition_penalty (many don't implement it), you may need to reset it to 1.0 or drop it. Multi-image support is endpoint-dependent too - a server that only handles one image per request will error when image2 is connected. As with all these API nodes, remember what leaves the machine: if the endpoint is a remote provider, your images go with your question.