JZL - 🌐 API 设置

"JZL - 🌐 API 设置" is the cloud-LLM counterpart to the pack's "模型加载Pro". The Script Processor that writes H3 storyboards can run on a local GGUF via JZL_LlamaModelLoaderPro, or it can call an API - and this node is the settings object for that path. You wire its single output into the Script Processor's api_config input and it tells the processor which provider, model, key, and sampling knobs to use.

The whole point of this node is that your API key never ends up inside the workflow. It's a single boolean open_settings button - click it and a dialog pops up, you pick a provider and paste the key, and on save the config is written to jzl_minimax_h3_api.json in ComfyUI's user directory with the key masked. The workflow itself only ever carries a reference, so sharing a .json workflow doesn't leak credentials. That's a small mercy that a surprising number of LLM node packs skip.

How it works

Three provider families are offered in the dialog: OpenAI-compatible (which covers OpenAI, DeepSeek, Qwen, GLM, Kimi, Ollama, vLLM, LM Studio - anything speaking the OpenAI schema), Anthropic, and Google Gemini. You configure the base URL, model name, temperature, and token limits there. At execution the node emits that config as a JSON api_config STRING, and the Script Processor reads it to make its calls. The config is loaded fresh each run, so editing the dialog mid-workflow takes effect without touching any node.

The inputs that matter

There's exactly one: open_settings - the button. The tooltip is the honest description: click it, configure provider/model/API key/address/temperature/tokens in the popup, save, done.

The single output, api_config, feeds the "剧本与镜头处理器" (Script Processor)'s api_config input.

When to use it vs. the local loader

If you have a decent GPU and a Qwen3.5-9B-class GGUF, the local path is free and private. If you're on a weak card, or the Script Processor needs a bigger model than your VRAM holds, the API path gets you there with one node. There's no strong reason to wire both; pick the path that matches your hardware. Note this is the LLM-for-writing chain, separate from H3 video generation itself - H3's weights still have to come from the official ComfyUI loaders.

How to install it

Same pack as everything else - ComfyUI Manager, search "ComfyUI-JZL-MiniMax-H3", or:

cd ComfyUI/custom_nodes
git clone https://github.com/wjluoxiao/ComfyUI-JZL-MiniMax-H3

Restart ComfyUI. The node needs no special dependencies - the dialog is front-end JS and the output is JSON.

Common issues

The usual snag is pointing an OpenAI-compatible provider at the wrong base URL or omitting /v1 where the provider requires it; the error surfaces as a 404 from the Script Processor, not from this node. Also, this pack is Chinese-first, so the dialog labels are bilingual but lean Chinese - if a provider isn't listed, pick the OpenAI-compatible family and fill in the model name yourself. The key stays out of the workflow, but it's still stored in plaintext JSON on disk, so lock down your ComfyUI user directory like you would any other credential file.