AI_ModelScope_image

Read the name and you'd assume this generates images. It doesn't - it reads them. AI_ModelScope_image is a vision-language chat node against models hosted on Alibaba's ModelScope, the model hub behind Qwen, Wan, and Z-Image (Alibaba's Tongyi umbrella has quietly become the community's default open-model source since 2025, and ModelScope is where they host and serve a lot of it, including inference APIs for their own models). Feed it up to four images and a question, get text back. If you actually want ModelScope to make an image, that's AI_ModelScopeT2I or AI_ModelScopeImageEdit - this one's for description, analysis, and Q&A about images you already have.

One thing worth calling out plainly: this node's actual input labels are in Chinese, not English - 预设模型 (preset model), 提示词预设 (prompt preset), 自定义系统提示词 (custom system prompt), 用户消息 (user message), 温度 (temperature), 最大长度 (max length), 超时时间 (timeout), and the image slots 图像1–图像4. That's not a display bug, it's how the node ships - the pack is Chinese-authored and this particular node didn't get an English localization pass the way some of the others did. Once you know what each label maps to, it's the same shape as the pack's other chat nodes.

The default model is Qwen/Qwen3-VL-8B-Instruct, with Qwen3-VL-235B-A22B-Instruct (bigger, slower, presumably better) and Qwen3-VL-8B-Thinking (a reasoning variant) as alternatives, plus two oddballs - iic/GUI-Owl-1.5-8B-Instruct and its Think variant - which are GUI-agent vision models trained to understand screenshots and interfaces rather than general photos, worth trying specifically if your use case is UI/screenshot analysis rather than photography. Top-P and 温度/temperature (0–2, default 0.7) control sampling the way they do on any LLM chat node; 最大长度/max length (1–32768, default 2048) is your response token budget; 超时时间/timeout (1–600 seconds, default 180) is how long the node waits before giving up on a slow response, worth raising if you're hitting the larger 235B model. seed and MODELSCOPE_API_TOKEN round out the required fields, plus 自定义模型/custom model as a free-text override. Outputs: AI回复 (the reply text) and status.

Installing it: through ComfyUI Manager (search "ComfyUI-Apt_Preset") or manually - cd ComfyUI/custom_nodes && git clone https://github.com/cardenluo/ComfyUI-Apt_Preset - then install.bat and restart. No local model weights needed; it's calling ModelScope's hosted inference.

Getting a token: create a ModelScope account at modelscope.cn and generate an API token from your account/profile settings - ModelScope's own site has decent free-tier quota for a lot of hosted models, which is a real point in its favor next to some of the other API nodes in this pack that gate almost everything behind a paid tier.

Troubleshooting: a blank or error AI回复 usually traces back to MODELSCOPE_API_TOKEN - check it's filled in correctly first. If a request against the big 235B-A22B model is timing out, bump 超时时间/timeout up from the 180s default rather than assuming the model or your token is broken; larger models genuinely take longer to respond. And if the response reads like the model ignored your images, double check you actually wired something into 图像1 and not a later slot - the schema treats the image inputs as ordered, not interchangeable.