LLM Text Generator

LLM Text Generator is the flagship of the zyd232 pack, and it's a lot more than a chat widget bolted onto a canvas. It's a complete OpenAI-compatible client in one node - point it at Ollama, vLLM, llama.cpp, or any real API, and it'll call {base_url}/v1/chat/completions, stream the answer to a floating panel in real time, and hand you clean text (and optionally the model's reasoning) as outputs your workflow can use. If you've ever wanted an LLM to rewrite your rough idea into a prompt, caption an image, or write dialogue for an image-to-video clip inside the same graph that generates, this is the shape of thing you want.

The name is almost a lie in the best way: for local servers like Ollama the default api_key of sk-no-key-required is all you need. No key, no account. You're only adding an API key if you point it at a hosted service.

What you actually set

The node has a wall of inputs, but a beginner only touches a handful:

• base_url - the server address (default http://127.0.0.1:8080). Swap it for whatever yours listens on; Ollama's OpenAI endpoint is usually http://127.0.0.1:11434, vLLM typically http://localhost:8000. The node appends /v1/chat/completions for you.
• model / model_NoVision - the vision model (used when you feed it images/video/audio) and the text-only model (used when you don't). You can type names or use the dropdowns; there's a refresh button that asks the server what it actually has.
• system_prompt and user_prompt - the usual suspects.
• temperature - higher is looser, lower is more stable.

Then the extras that earn their keep: config_name + the 💾 save button stores everything (URL, key, models, prompts) as a named preset and hides your API key behind ********; the key also accepts ENV:MY_VAR so you never paste a secret into a workflow you share.

Multimodal, streaming, and the output that matters

This is where it stops being a toy. The images, videos, video_audios, and audios inputs are autogrow - connect one, another appears, up to 32 each. Images go in as base64 PNG; videos get frame-sampled (the video_fps control samples frames from a 24fps source, capped by max_video_frames so you don't blow the context window); audio needs enable_audio flipped on and a model that supports it.

The node streams tokens over ComfyUI's websocket to a floating "Streaming Text" panel that follows the node around. Two outputs come out the right side: text (the answer) and reasoning (the thinking process, if you turn on thinking - it wraps the model's chain-of-thought in <think>/</think> tags and separates it from the final answer).

The one feature worth calling out: Lock result. Run once, hit 🔒, and the output is embedded in the workflow. Re-run (or send the file to a friend) and the node returns the locked text without calling the LLM at all - no API cost, no dependency on the server still being up.

Install

This is the only "heavy" node in the pack, and even it is light. The real dependency is a running LLM server - the node needs no OpenAI SDK, just Python's stdlib plus aiohttp, which ComfyUI already ships. Install as usual:

cd ComfyUI/custom_nodes
git clone https://github.com/zyd232/ComfyUI-zyd232-Nodes.git

Restart, and search "LLM Text Generator" (category zyd232 Nodes/LLM). Note the pack is built on the V3 backend node API, so keep ComfyUI reasonably current.

Where people get burned

• Wrong base_url. The default 8080 is nobody's real default; a "connection refused" almost always means the server isn't there or isn't running.
• Dropdowns are empty until you refresh. Hit 🔄 Refresh Model List once the server is up.
• Numbering off by one. Input ports are image_0, image_1… but the prompt labels sent to the model are 1-based - refer to image_1 in your prompt when you mean port image_0. It's a documented quirk, but it'll trip you up exactly once.
• Fallbacks exist. A 404 on /v1 auto-retries without it; a failed no-vision model falls back to the vision model. Usually invisible, occasionally surprising.
• It's a small, new pack. LLM nodes are arbitrary Python that reaches the network - the exact category weaponized once already (the LLMVISION incident). Read what a fresh clone does before trusting it with a real API key; this one's source is open - do that.

Bottom line

If you already run Ollama and want a clean, streaming, multimodal LLM node inside your workflow - presets, reasoning separation, lock-result escape hatch - this is a strong pick.