ZhiPu Ai Node (fsdymy)

This is the pack's most interesting node, and also the one with the most setup attached. Instead of running a local LLM on your GPU, it calls Zhipu AI's (智谱AI, bigmodel.cn) cloud chat API - the GLM family - to expand and rewrite your prompt. That's the "LLM-assisted prompting" pattern the community converged on: an LLM that understands language writes the instruction, your image model renders it. It went from a browser-tab trick to a whole node category because it fits the architecture rather than fighting it. This node is the Chinese-market, no-local-weights version of that idea: you pay Zhipu per call, you need no GPU headroom for the LLM part.

How it works

Read the source and the shape is clear. The node keeps a session_history list in memory on the node instance. Each run it builds a message list: your system_content first, then the last context_size turns of the conversation, then your new prompt. It sends that to Zhipu's chat.completions.create, gets the reply, and appends both your prompt and the reply to the history. context_size (0–30, default 1) crops how much memory is kept - set it to 0 and every call is stateless, which is what you want for one-shot prompt expansion.

The inputs that matter

• prompt - your rough idea ("1girl" is the default). Multiline.
• system_content - the system prompt, and the real lever. The default is a Chinese "creative dream-painter" assistant instructed to answer in English. Change this to control output format: for a tag-based model like Illustrious you want it told to emit comma-separated booru tags; for an LLM-encoded model you want structured prose. This is where you do the steering.
• model - defaults to glm-4; the list also has glm-4-flash, glm-4-air, glm-4v, and several 2024-era variants. glm-4-flash is the cheap/free-tier chat model and is the sane default for this job. Don't pick embedding-2 or charglm-3 - they're not chat-completion models and the call will fail.
• context_size - how many prior turns of memory to keep, default 1.
• seed - present in the schema, but here's the honest part: the code never actually uses it. It's a vestigial input left over from an earlier design. Any value does the same thing.

Outputs

Three strings. text is the expanded prompt - that's the one you care about; wire it into a text-to-prompt path or just copy it into your CLIP encode. messages is a JSON dump of everything sent in this call, handy for debugging. session_history is the running conversation as JSON.

Setup, because there is setup

You need a Zhipu account and API key from bigmodel.cn. Then edit the pack's key.json - the shipped default key is literally "xxx", so nothing works until you change it:

{ "url": "https://open.bigmodel.cn/api/paas/v4", "key": "your-real-key-here" }

It lives at ComfyUI/custom_nodes/ComfyUI_fsdymy/key.json. Install the pack with ComfyUI Manager (search "ComfyUI_fsdymy") or git clone https://github.com/fsdymy1024/ComfyUI_fsdymy, and its requirements.txt installs the zhipuai SDK. Then restart.

Where people get burned

• No key or a bad key → the node errors on the first run. That's not a bug; there's no UI for the key, only the file.
• It needs internet. Your prompt text leaves your machine for Zhipu's API. Fine for most people, a hard no if you work offline or under strict data policies.
• Session memory lingers. Because history lives on the node instance, the conversation accumulates across runs until you reload the workflow. That can be a feature (multi-turn refinement) or a footgun when it quietly changes what the model thinks you meant. context_size: 0 kills it.
• The model list is frozen in 2024. Zhipu has moved on; the node's list hasn't. glm-4 still works, but expect the bleeding edge to be missing.

The pack itself is a single-commit, unmaintained hobby repo, and this node is the part most likely to drift as the API changes. For one-off prompt expansion it's genuinely handy. For production workflows, the key-file setup and stale model list make me reach for a more maintained LLM node - but if you're already a bigmodel.cn customer and want the Chinese-market path, this works.