AI_ModelScopeT2I

This is the straightforward text-to-image node of the ModelScope group - no reference images, no editing, just a prompt in and a picture out, hosted on Alibaba's ModelScope inference API. The default model is telling: Tongyi-MAI/Z-Image-Turbo, which is one of the more genuinely impressive open releases of the last year - a 6B model that runs comfortably on a 12GB card, roughly 30 seconds per image, uncensored, Apache 2.0, and quick enough on release that a chunk of the community started calling it "SDXL's successor" once the trainable Base variant landed. If you already run Z-Image locally, this node buys you nothing over your own GPU - but if you don't have local VRAM to spare, it's the same model as a paid-per-call API instead. The other two dropdown options are Qwen/Qwen-Image-2512 and MoYouuu/MYHuman-QWEN, a community fine-tune.

Required fields, all with the Chinese labels this pack's ModelScope nodes ship with: 预设模型 (preset model), 提示词 (prompt - the field ships with the author's own placeholder default, "a cute girl in festive chinese new year clothing," which you'll obviously want to replace), 图像宽度/图像高度 (width/height, 64–8192 in steps of 64, default 1024×1024), 张数 (image count, 1–4), 超时时间 (timeout, 1–900 seconds, default 300), seed, MODELSCOPE_API_TOKEN, and 自定义模型 (custom model override). No optional image inputs at all - this node is pure T2I, which is also how you can tell it apart from AI_ModelScope_image (which despite its similar name is actually an image-understanding chat node, not a generator). Outputs: 图像 (the generated IMAGE), 图片链接 (a direct URL to the result), and status.

Because Z-Image-Turbo is small and fast by design, this is one of the more forgiving nodes in the AI-tool set to experiment with - you're not waiting on a 20B+ parameter model, so 超时时间's generous 300s default should rarely get exercised at the Turbo default, though the larger Qwen-Image-2512 option will eat more of that budget.

Installing it: ComfyUI Manager, search "ComfyUI-Apt_Preset," or clone by hand - cd ComfyUI/custom_nodes && git clone https://github.com/cardenluo/ComfyUI-Apt_Preset - then install.bat and restart.

Getting a token: register at modelscope.cn and generate an API token in your profile. ModelScope's free quota is genuinely usable for testing this node without committing to a paid plan first.

Troubleshooting: an empty result or an auth error traces back to MODELSCOPE_API_TOKEN nine times out of ten - confirm it's current and pasted cleanly. If you're getting worse results than you'd expect from Z-Image's reputation, remember its known gaps are real and not a sign something's misconfigured: weaker adherence to very complex, multi-clause prompts than Flux 2, a persistent bias toward Asian faces in ambiguous prompts, and roughly a 2-megapixel native ceiling - pushing 图像宽度/图像高度 well past that ceiling is more likely to degrade quality than improve it, so if you need genuinely higher resolution, generate at the native size and upscale afterward instead.