Qwen视觉丨API

Sometimes the thing your workflow needs is someone to look at the image. QwenVisionAPI is the QING pack's VLM node: it sends an image plus a text question to Alibaba's Qwen-VL models over the API and returns the analysis as a string. Captioning, OCR, "what's wrong with this face," image-to-prompt - all without running a vision model locally. It's the same hosted-LLM pattern as the pack's QwenLanguageAPI, but with eyes.

How it works

Inputs: image (the IMAGE to analyze), text_input (your question, default "describe this image"), platform (阿里云百炼 / 硅基流动), model (qwen2.5-vl-72b-instruct, qwen3-vl-plus, qwen-vl-max-latest, or qwen3-vl-235b-a22b-instruct), max_tokens (default 2048), and history (default 5 - vision conversations are usually short). Optionals: temperature (0.3), top_p (0.8), image_quality (auto / low / high), and clear_history.

Outputs: analysis_result (the model's text answer), conversation_info, and total_tokens.

Same provider plumbing as the language node - OpenAI-compatible endpoints on DashScope or SiliconFlow, keys from env (DASHSCOPE_API_KEY / SILICONFLOW_API_KEY) or the QING settings panel, and model names mapped to real API IDs. image_quality is the knob worth knowing: low sends a downscaled image for speed (fine for OCR and broad description), high sends more detail for fine visual reasoning, auto lets the API decide. Conversations cache per platform+model so you can hold a follow-up dialogue about the same image.

Where you'd reach for it

Captioning is the big one - feeding a LoRA training run with consistent natural-language captions is exactly the VLM-in-the-loop job from the KB's lora-training doc, and Qwen-VL is strong at it (including, importantly for this ecosystem, uncensored coverage). It's also the cleanest way to do image-to-structured-prompt: describe a reference image, feed the description to a prompt template, generate a variation. In a batch pipeline, pair analysis_result with a text compare or QING_ModelFallback to gate downstream work on what the model saw.

Installing and configuring

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

Restart ComfyUI after. The openai client (pulled by the installer) is the only real dependency. Then get a key - DashScope at bailian.console.aliyun.com or SiliconFlow at cloud.siliconflow.cn - and enter it under Settings → QING → API configuration, or set the env var. Keys stay local and are injected at runtime, never embedded in your workflow. (Ignore the README's GAOSHI-QING clone-URL typo; the repo is GAO-SHIQING/ComfyUI-QING.)

Things to know

Token math is where vision bills add up: every call ships the image as base64 plus your text, so max_tokens and history compound quickly - that's why the defaults are smaller than the language node's. Use image_quality = low for high-volume jobs where precision doesn't matter. Network and key required, obviously, and a SiliconFlow account only exposes the open qwen2.5-vl model while DashScope has the whole roster. And like any freeform VLM output, validate before you trust: for OCR or caption work the result is usually excellent, but for programmatic branching you'll want to constrain the question hard.