EA LM Studio

You already know the workflow: you're building a prompt-enhancement chain for Flux or Qwen-Image, you want a VLM to describe an image into a prompt, or you just want an LLM to draft options you then steer by hand. You also already have LM Studio sitting on your machine with a Qwen3 or DeepSeek loaded. EA LM Studio is the bridge - a single node that turns ComfyUI into a chat client for whatever model is loaded in your local LM Studio server. No cloud API, no key, no per-token billing. The model stays in your machine's RAM.

The name trips people up once: this node does call an "API," but it's LM Studio's own local server on 127.0.0.1:1234 - nothing leaves your box. That's exactly the right posture for an LLM node, given how the ecosystem got burned. The ComfyUI_LLMVISION debacle (an LLM-vision node that shipped infostealer wheels and eventually landed its author 15 months federal) is the reason to prefer open-source nodes like this one that only talk to your own server. Read the code if you're nervous; it's small enough.

How it works

Under the hood it uses the official lmstudio Python SDK to talk to the local server. Models are discovered automatically at ComfyUI startup - the dropdown populates from LM Studio's model list, so you pick by name instead of typing a path. A refresh_models toggle re-fetches instantly if you loaded a new model after ComfyUI started.

The clever bit is reasoning extraction. For thinking models like DeepSeek R1, Qwen3, and QwQ, the node splits <think>...</think> content off into a separate reasoning output, so your downstream nodes never choke on chain-of-thought. It even handles GPT-OSS's "harmony" channel markers (<|channel|>, <|message|>, <|end|>, <|return|>) and cleans up partially-leaked tags.

The inputs that matter

Most of the node is boilerplate sampling knobs you can ignore. You actually set these:

• system_message and prompt - your role definition and your request. The two you'll touch every run.
• model_selection / custom_model_name - pick a discovered model, or choose "Custom" and type an identifier from LM Studio's /v1/models list.
• max_tokens - output length only; the input side is governed by the context window you set in LM Studio itself.
• temperature - 0.7 default, drop toward 0.1 for deterministic work.

Two footguns worth knowing. seed looks like it should make runs reproducible - it can't. The LM Studio SDK has no inference-time seeding, so the seed is purely a re-roll control: changing it tells ComfyUI the node is dirty and should run again. Don't chase reproducibility through it. And if output turns repetitive, reach for repeat_penalty (1.1–1.3) or presence_penalty (0.3–0.8); min_p at 0.05–0.1 is the modern, less finicky alternative to top_p.

For vision models, wire image1 through image4 and set image_resize - the auto-resize exists precisely to stop multi-image VLMs from OOMing. For Qwen3 and friends, enable_thinking forces thinking on hybrid models without the /think prompt hack, and reasoning_mode controls how thinking gets split out. unload_llm (on by default) frees VRAM after generation so your diffusion models can load back.

Outputs

Three strings: response (the cleaned answer), reasoning (the extracted thinking), and troubleshooting - status messages plus real stats like tokens/sec, time to first token, and stop reason. Wire response into a text display or straight into another node's prompt input; that's the whole point.

Installing it

ComfyUI Manager: search "EA_LMStudio". Or the manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/EnragedAntelope/EA_LMStudio.git
pip install -r EA_LMStudio/requirements.txt

Dependencies are light - the lmstudio SDK, requests, Pillow, numpy - but the SDK needs Python 3.10+, and you need LM Studio running with its server enabled (validated against 0.4.17). Server host/port and model-exclusion patterns live in lms_config/user_config.json.

Common issues

Models not showing? LM Studio must be running before ComfyUI starts, then toggle refresh_models. Context errors? Raise the context length in LM Studio settings, not max_tokens. enable_thinking doing nothing? The toggle sets an enableThinking flag that only works when the model's chat template honors it - many community finetunes ignore it, so fix it in LM Studio by editing the Jinja template or its reasoning-parsing delimiters. VLM failing on multi-image? Drop to one image or a smaller image_resize. And keep both LM Studio and the SDK updated: older backends silently ignore newer sampling params like min_p rather than erroring, so stale versions quietly no-op half your settings.