💻 VRGDG Local LLM 💻

The name might make you reach for a credit card, but VRGDG_LocalLLM runs entirely on your machine by default. It's a text-generation node that talks to a local LLM server - Ollama out of the box - and rewrites or enhances a prompt for you. No API key, no cloud bill, no "your prompt was logged somewhere" dread. The key box exists because the node can also talk to any OpenAI-compatible server, but the default path is fully local.

This is the pack's prompt-enhancer brain. Modern image and video models are picky about prose - Z-Image wants natural language, and LTX-2 famously shipped a Gemma-based prompt enhancer because these models reward long, detailed prompts over terse tags. A local LLM that beefs up your "woman in a city at night" into a full cinematic paragraph is the same job, done your way, offline.

How it works

It's an HTTP client with a good default. backend picks ollama (default) or openai_compatible. With Ollama, base_url points at http://127.0.0.1:11434 and model names a model you've already pulled - the default is qwen2.5:14b, which you'd grab with ollama pull qwen2.5:14b. With an OpenAI-compatible server (LM Studio, vLLM, a hosted endpoint), api_key and custom_model come into play.

The inputs that matter:

• prompt and system_prompt - your task and the persona ("You are an expert prompt engineer" is the default). This is where you actually steer output.
• temperature - 0.4 default. Low for consistent rewrites, higher if you want variety. This node is used for batch prompt generation, so keep it low.
• max_tokens - 4096 default, plenty for prompt enhancement.
• image1 through image4 (optional) - if your local model is vision-capable, you can hand it reference frames and have it describe or build on them. A real workflow for the music-video pipeline: show it the character reference, get a prompt back.

Outputs: text (the generated prompt - wire this into your prompt encoder), plus used_backend, used_model, and status for debugging, and an image output that passes a reference image through if you connected one.

The trap in the category name

The category is VRGDG/NanoBananaPro, which tells you where this node actually lives: inside NanoBanana workflows. NanoBanana (Google's image model) has its own prompt-enhancement step, and this node slots in as a replacement or alternative so you control the wording. You don't need to know NanoBanana to use the node - but if you pulled this out of a workflow that references it, that's why it's grouped there.

Getting it and troubleshooting

Same shared pack:

cd ComfyUI/custom_nodes
git clone https://github.com/vrgamegirl19/comfyui-vrgamedevgirl

or Manager → search "vrgamedev", restart.

The usual failure modes are all "is the server running" problems: Ollama isn't started, the model name is mistyped (qwen2.5:14b vs qwen2.5:7b), or the port changed. request_timeout_sec (default 300) exists because a 14B model on CPU can take a while - if it times out, raise that, not your temperature. And if you get empty output on a first run, check that you actually pulled the model: ollama list is your friend. Once it's up, you've got a genuinely useful prompt-engineering workhorse that costs you exactly nothing per call.