LM Studio Model

This is the node you drop first, because every other node in the WAS LMStudio Easy-Query pack wants its model input wired to what this one outputs. If you're used to ComfyUI loaders that yank weights into your own VRAM, adjust your expectations: LM Studio Model doesn't load anything. It points the rest of the pack at a model that's already running inside the LM Studio app on your machine, and bundles the sampling defaults for that model into one wire.

The name is a small lie. There's no API key, no cloud call, no huggingface.co download here. This pack talks to LM Studio through its official Python SDK, over localhost, to a server you're running on the same box. That's the whole trick that makes a vision-language model usable from inside ComfyUI without fighting it for VRAM - LM Studio owns the model, ComfyUI just sends prompts and gets text back. It's the same pattern the community has been wiring together by hand for a while now: a Qwen3-VL-class model captioning images so you can feed the results to Z-Image or Flux or whatever you're actually generating.

The inputs that matter:

• model - a dropdown populated at graph load by asking LM Studio's SDK which models you've downloaded. If it shows <no models found>, LM Studio isn't running or has nothing installed yet; the pack defaults to qwen/qwen2.5-vl-3b, a good small vision model to start with.
• manual_model_id - the escape hatch. If your instance isn't returning models via /models, or runs on a different base URL, type the identifier here (e.g. qwen/qwen2.5-vl-7b) and it overrides the dropdown.
• temperature / max_tokens / seed - your baseline sampling. Set max_tokens to -1 for unlimited, which the tooltip recommends for thinking models. seed 0 disables it.
• image_max_size - the longest edge images get downscaled to before they're encoded and shipped to LM Studio. 1024 is a sane default; drop to 512 if you're on a small context window.
• unload_after_use (default on) - unloads the model after each query. On a single GPU this is the setting that lets you caption a batch, then immediately generate images without an OOM.
• use_rest_api - flips the backend from the SDK to LM Studio's OpenAI-compatible REST endpoint. Required for proper thinking-model support, because it separates the reasoning text from the actual content.

The single output, model of type LMSTUDIO_MODEL, is a dict carrying model_id, temperature, max_tokens, seed, image size, and the unload flag - everything the Query, Caption, Chat, and dataset nodes need. Think of it as a config object, not a loaded model.

Install it the usual way: ComfyUI Manager, search "WAS LMStudio Easy-Query", or

cd ComfyUI/custom_nodes
git clone https://github.com/WASasquatch/ComfyUI_LMStudio_EasyQuery

then restart ComfyUI. The pack needs LM Studio's Python SDK plus numpy and Pillow:

pip install lmstudio numpy Pillow

and, separately, the LM Studio app running with at least one model downloaded inside it - the pack won't fetch models for you.

If the dropdown comes up empty, check that LM Studio is actually running and serving, then try manual_model_id. And remember the pack reads lmstudio_config.json in its own folder - you can set a default_model and the image_max_sizes list there, but you must restart ComfyUI after editing. Those SDK websocket shutdown messages you'll see in the console when ComfyUI closes are normal, not an error.