GLM视觉丨API

Captioning, image QA, "describe this and turn it into a better prompt" - that's the vision-model job, and GLMVisionAPI is the QING pack's route into Zhipu's GLM vision models. You feed it an image and a question, it returns a text answer, and like the pack's other chat nodes it keeps a short multi-turn history so you can interrogate the same image or refine a description across runs. If you're building the caption-image → prompt-enhancer → sampler chain that the KB's LLM essay describes as the mainstream 2026 pattern, this is a serviceable API node for the first link.

It's part of ComfyUI-QING's API category, sharing the pack's framework with the language nodes - so platform, key, and history handling all work the same way. The platform dropdown offers 智谱AI and 硅基流动, and the vision model roster differs between them.

How it works. The image input is the one required visual: your IMAGE tensor gets sent to the model along with text_input (default "请描述这张图片的内容" - describe this image). The platform routes to Zhipu or SiliconFlow; the model field lists each host's vision lineup. Zhipu side you get the mainstream GLM-4V, GLM-4V-Plus, GLM-4V-Flash, GLM-4.5V; SiliconFlow contributes GLM-4.1V-9B-Thinking (a 9B chain-of-thought model, and the default) and GLM-4.1V-Thinking-FlashX. The image_quality knob (auto/low/high) trades speed against detail, same as the other vision nodes in the pack.

The inputs that matter.

• image - what the model looks at.
• text_input - your question or captioning instruction.
• platform + model - host and model; keep the platform matched to the key you hold.
• max_tokens (default 4096) and history (default 8 rounds).
• Optional: temperature, top_p, image_quality, clear_history.

Outputs: generated_text, conversation_info, total_tokens. Wire generated_text into a text preview, a prompt template, or a filename builder.

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

Then configure the key in Settings → QING → API Configuration and restart. The README's clone URL is typo'd (GAOSHI-QING, missing the H) - use the URL above or Manager.

Troubleshooting. Same family of issues as the pack's other API nodes: platform/key mismatch is the top error (Zhipu dropdown + SiliconFlow key = 401), and some models exist on only one host. Vision calls are slower than language calls, especially with image_quality at high - if your workflow times out, drop the quality or the max tokens. And the usual API-node cautions apply: your image and question leave the machine to the provider, and the model's content policy applies - GLM vision won't describe what it's trained to refuse. For anything NSFW or attribution-sensitive, the local captioner path (JoyCaption, Florence-2) remains the community's default.