魔搭❤免费生图

The pack's Gemini nodes generate images through a paid reseller channel. This node is the budget lane: it runs Tongyi-MAI/Z-Image-Turbo - Alibaba's 6B image model, the one the KB's Z-Image panel calls the community's local-photorealism default - through Alibaba's ModelScope inference API instead of on your GPU.

That's a genuinely different trade than the Banana nodes. Z-Image is an open model (Apache 2.0) that you could run locally on a 12GB card; this node is the "no GPU, or don't want to host it" path to the same weights, per the KB's external-api-nodes framing of calling an open model on someone else's server. For batch work where you don't want to babysit VRAM, that's a reasonable exchange - and it uses a separate key and billing from the Gemini channel.

The inputs that matter

• prompt (multiline) - the positive prompt. Z-Image's Qwen3 text encoder handles both English and Chinese natively.
• modelscope_api_key - required. Get one from ModelScope (魔搭社区) - this is the same account/hub the KB's ModelScope panel covers, and the key is for its inference API.
• model - currently just Tongyi-MAI/Z-Image-Turbo, with a custom_model field to override with any ModelScope-hosted model name (leave blank to use the dropdown choice). If the author later adds options, they'll appear in the dropdown; custom_model is the escape hatch for everything else.
• negative_prompt - supported, which matters: the KB notes negative-prompt support is a real Z-Image feature.
• width/height (64–2048, step 64, default 1280×1280) - output size. Z-Image has a ~2MP native ceiling per the KB, so don't expect the 4K behavior of the Gemini Pro models; keep it under ~2048 per side.
• steps (1–100, default 10), guidance (1–20, default 1.5) - Turbo is a few-step model, and the defaults are the sensible starting point; 10 steps is plenty for Turbo.
• seed (-1 randomizes), batch_size (1–15), max_concurrency (1–10, default 4) - the batch size is generous and the concurrency knob parallelizes the requests; keep max_concurrency modest unless you know the endpoint can take it.
• 绕过代理 and 禁用SSL验证 - the pack's usual network toggles.

The one output

images (IMAGE) - the batch of generated images, ready for a preview or save node. Batch count follows batch_size.

Installing and gotchas

Same pack as the rest:

cd ComfyUI/custom_nodes
git clone https://github.com/98624017/comfyui-banana-li-linux

Restart ComfyUI. No weights to download - the model runs on Alibaba's side.

Three things to know before you burn credits. First, key discipline: this uses the ModelScope key, not the Banana one, and the pack's global key node (XinbaoApiKeyPurge) has a dedicated modelscope_global_api_key slot for exactly this. Second, because Z-Image is open, remember the local alternative exists - if you have a 12GB+ card, running Z-Image Turbo locally (it's ~30s/image and uncensored) costs electricity instead of API calls, and you might prefer that for high volume. This node is the right call when you don't have the card or want the batch to not touch your GPU. Third, Z-Image's biases are real (an Asian-face bias, per the KB) - if you're generating diverse people, feed it explicit descriptions rather than leaving faces to chance.