LM Studio Query

Want an LLM inside your workflow without paying for an API? LM Studio Query talks to a local LM Studio server - which is just your own GPU running an OpenAI-compatible endpoint at http://localhost:1234 - and can do two genuinely useful things. First, pure text generation: feed it a template prompt and it writes a finished prompt for your sampler. Second, vision: attach images and it describes them, which makes it a batch captioner for training data or a "what did it actually generate" inspector. No key, no cloud, no monthly bill.

How it works

The node hits LM Studio's /v1/models endpoint to list what's loaded and available, then sends your system_prompt and user_prompt to the chat completions endpoint. If you give model_name (full or partial), it loads that model first; leave it blank and it uses whatever's currently loaded. Images get base64-encoded and sent along, so a vision-capable model can describe them. prompts_number runs the query multiple times with different seeds - handy for getting several prompt drafts to pick from - and unload_after_use frees the model from memory when done.

The inputs that matter

• system_prompt / user_prompt - the usual LLM pair. The defaults ("You are a helpful assistant." / "Describe this image in detail.") are fine starting points.
• server_url - http://localhost:1234 unless you changed LM Studio's port.
• model_name - leave empty for the loaded model; otherwise a name or substring to load.
• temperature, top_p, max_tokens, seed - standard sampling knobs. 0.7 / 0.9 / 256 defaults are sensible.
• prompts_number - how many generations (1–10).
• images - optional IMAGE input; supports batches, so you can caption several images in one run.

Outputs: generated_texts (the responses, as a list if you asked for more than one) and model_ref (the model that actually ran - handy if you want to feed it into the unload node).

Installing it

Install the pack, and make sure LM Studio is actually serving:

cd ComfyUI/custom_nodes
git clone https://github.com/Pirog17000/Pirogs-Nodes
pip install -r Pirogs-Nodes/requirements.txt

In LM Studio: enable the local server (Settings → Server, default port 1234) and load a model. The lmstudio package in requirements.txt is technically only needed for unloading; the query node itself just needs requests.

Gotchas

Most failures here are "LM Studio isn't actually running" - the node errors out if it can't reach /v1/models in a couple of seconds. If you specified a model name and it's not in the server's list, you get a "not found" error. Vision queries silently degrade if your model isn't multimodal, so check which model you loaded. And remember the output is text: you still need to wire generated_texts into your prompt or a text-encode node - this node writes prompts, it doesn't apply them.