ModelScope-Image 生图节点

This is the flagship of the ModelScope pack: a plain text-to-image node that turns a prompt into a real IMAGE tensor you can preview, save, or pass on to an upscaler. What makes it interesting isn't the node - it's what's behind it. The default model is Qwen/Qwen-Image, a 20B MMDiT that would otherwise want a big chunk of VRAM and a multi-gigabyte download. Here the actual diffusion happens on ModelScope's cloud, so you can generate with Qwen-Image, a MusePublic FLUX.1-Kontext-Dev, or Z-Image-Turbo with zero local model files and (in theory) a 6GB card. That's the real pitch: pick any model from the dropdown, pay per render, never download weights.

How it works

Mechanically it's more interesting than most cloud nodes. It POSTs to https://api-inference.modelscope.cn/v1/images/generations with an X-ModelScope-Async-Mode header, gets back a task_id, then polls /v1/tasks/{task_id} every five seconds until the job reports SUCCEED, downloads the image, and converts it to a ComfyUI tensor. Async because Qwen-Image is slow, even on their GPUs; the config file sets a default 720-second timeout. The console is chatty through all of it - every step, status, and seed is logged with emoji, which is either endearing or exhausting depending on how long you've watched a task poll.

The inputs that matter

Inputs you'll actually set:

• prompt and api_tokens (both required) - the prompt is self-explanatory; the token you get from modelscope.cn. You can paste several tokens separated by commas or newlines, and the node falls through them if one hits a rate limit. First run auto-saves them, so later runs can leave the field blank.
• model - dropdown default is Qwen/Qwen-Image, with Qwen-Image-2512, MusePublic/489_ckpt_FLUX_1, MAILAND/majicflus_v1, Tongyi-MAI/Z-Image-Turbo, and more. The list is read from modelscope_config.json, so you can edit that file to add models.
• negative_prompt - only sent if you actually type something. Qwen-Image is picky about negatives; keep them short.
• width/height (64–2048, step 64), steps (1–100), guidance (1.5–20, default 7.5) - the usual knobs. Default is 512×512, which is conservative; most people push 1024.
• seed (-1 = random) - set it to reproduce a render, exactly like a local sampler.
• lora1_id / lora1_w through lora3_id / lora3_w - hosted LoRAs, by ModelScope model ID, sent as a loras dict in the request. This is where the pack's LoRA loader nodes plug in.

Output is a single image - wire it into PreviewImage or SaveImage.

Installing it

Install is the standard two ways: ComfyUI Manager, searching "ComfyUI ModelScope API Node", or:

cd ComfyUI/custom_nodes
git clone https://github.com/hujuying/ComfyUI-ModelScope-API

then restart. Dependencies are light - everything in requirements.txt (requests, pillow, torch, numpy) already ships with ComfyUI; only the text/vision nodes need the openai package.

The honest caveats

The honest caveats: this is a paid API, not free inference - your free ModelScope quota evaporates fast if you batch. And you're at the mercy of their queue: a render is a network round trip plus polling, so "30 steps" takes as long as their GPUs say it takes, not as long as your GPU would. It's a great node when you're on a weak card or testing models you can't run; it's a strange pick when you own a 4090 and the model already lives in your models folder.