LM Studio Options

LM Studio Options is the fine-control node in the WAS LMStudio Easy-Query pack, and it does exactly one job: hand a set of sampling overrides to whichever request node it's plugged into. The pitch is simple - your Model node sets your baseline temperature, max_tokens, and seed, and this node lets you override them for a single request without reconfiguring the whole graph.

Why does that exist? Because in a real workflow you don't want the same generation settings for every call. A captioning pass wants low, boring, deterministic temperature so your .txt files come out clean; a "rewrite this prompt" step wants more randomness so you get variety; a chat about your workflow might want a long output cap. The Model node is one shared configuration object - Options is how you branch. Anything it provides takes precedence over the Model node's values, per the tooltips.

Inputs worth knowing:

• temperature (default 0.2) - the usual knob. Lower = more deterministic, and for captioning you almost always want it low.
• max_tokens (default 512) - cap on new tokens per response; -1 means unlimited, recommended for thinking models.
• seed - 0 disables; set a nonzero value if you want reproducible sampling where LM Studio supports it.
• top_p (default 1.0, meaning off) and top_k (default 0, meaning off) - the classic nucleus/top-k filters, only if you're chasing a specific feel.
• frequency_penalty, presence_penalty, repeat_penalty - repetition control. Frequency penalizes tokens proportionally to how often they appear; presence penalizes them just for appearing at all; repeat_penalty is the generic engine-specific one (1.0 = off). If your captions come back looping on the same words, this is the section to touch.
• strip_thinking_tags (default off) - the one that actually saves you time. It strips <thinking>...</thinking> / <think>...</think> blocks out of the output. A thinking model without this will happily dump its entire chain-of-thought into your caption file.
• stop (optional) - comma-separated strings that halt generation when any of them appears. Leave blank for none.

The output is a single options wire of type LMSTUDIO_OPTIONS. It plugs into the optional options socket on the Query, Caption, Caption Dataset, and Chat nodes. That's it - the node does no model work on its own, which is why you'll always see it sitting upstream of a request node rather than standalone.

Installation is the same as the rest of the pack - it ships inside WASasquatch/ComfyUI_LMStudio_EasyQuery, so if you have that installed you already have this node. Via Manager, search "WAS LMStudio Easy-Query":

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

Restart ComfyUI, and it needs the lmstudio SDK installed (pip install lmstudio numpy Pillow) plus LM Studio running locally - though strictly, Options itself only builds a dict; it's the request node downstream that needs the server alive.

One gotcha worth naming: the seed here overrides the Model node's seed, so if you want reproducibility across an Options-fed request, set it here, not on the Model node. And if you're captioning a dataset with a thinking model, flip strip_thinking_tags on early - scrubbing reasoning text out of hundreds of caption files after the fact is a chore nobody wants.