ComfyUI LM Studio Nodes

Use a local LM Studio server as the LLM/VLM backend inside ComfyUI — a drop-in replacement for the built-in CLIPLoader → TextGenerate pair, useful for local image captioning (e.g. building LoRA training datasets) or general text generation, with no API keys or cloud calls required.

Nodes

LM Studio Connection

Replaces CLIPLoader. Points at your running LM Studio server.

Input	Description
base_url	LM Studio's OpenAI-compatible endpoint. Default: http://localhost:1234/v1
model	Exact model id as shown in LM Studio's Developer tab (e.g. gemma-4-e4b-it-qat). Leave blank to use whatever model is currently loaded.
api_key	Ignored by LM Studio, but must be non-empty. Default lm-studio is fine.
timeout	Request timeout in seconds (default 120).

On connect, it pings /v1/models immediately so you get a clear error right away if the server isn't running or the model name doesn't match, instead of a failure mid-batch.

LM Studio Generate Text

Replaces TextGenerate. Sends a chat completion request to LM Studio.

Input	Description
connection	Output from LM Studio Connection.
prompt	Your instruction / captioning prompt.
max_tokens	See "Minimum tokens" below — this is the setting most likely to bite you.
temperature, top_p	Standard sampling controls.
seed	Set to 0 for no fixed seed, or a specific value for reproducibility.
image (optional)	Wire in an IMAGE output (e.g. from LoadImage) for vision captioning. Leave unplugged for a plain text assistant.
system_prompt (optional)	Persistent instruction/persona sent as the system message.

Output generated_text is a plain STRING — wire it into ShowText|pysssss, SaveText|pysssss, or anywhere else a STRING input is expected.

Setup

1. Drop this folder into ComfyUI/custom_nodes/ (folder name doesn't matter, just don't nest it inside itself).
2. Restart ComfyUI. The two nodes appear under the LM Studio category.
3. In LM Studio: load your model, then go to Developer → Start Server. Confirm the port shown there matches base_url in the connection node (default 1234).
4. In ComfyUI: add LM Studio Connection → LM Studio Generate Text, wire an image in if you want captioning, and connect generated_text downstream.

Minimum tokens

Set max_tokens to at least 2000, and 8000 if your model uses "thinking"/reasoning mode.

Some models (reasoning-tuned Gemma, Qwen "thinking" variants, DeepSeek-R1 distills, etc.) write out a long internal reasoning trace into a separate reasoning_content field before producing the final answer in content. If max_tokens runs out mid-reasoning, LM Studio returns finish_reason: "length" with an empty content field — the node will fall back to using the raw reasoning text so you still get something, but it's messy chain-of-thought, not a clean caption, and you'll see a warning printed in the ComfyUI console when this happens.

Guidance by model type:

• Non-reasoning models (e.g. plain Gemma 3/4 IT, most vision-instruct models without a "thinking" toggle): 500–1000 tokens is typically enough for a one-to-two sentence caption.
• Reasoning/thinking models: 4000–8000 tokens. These models can burn several hundred to a few thousand tokens reasoning before writing a final answer, and this varies per image — set the ceiling generously rather than tuning it tightly.
• If you want speed over headroom, check whether your model exposes a way to disable thinking mode (a toggle in LM Studio's model settings, or a chat-template flag) — running without reasoning is both faster and keeps max_tokens low and predictable.

Sample workflow

LoadImage ──────────────┐
                         ▼
LMStudioConnection ──▶ LMStudioGenerateText ──▶ generated_text ──┬──▶ SaveText|pysssss
                                                                  └──▶ ShowText|pysssss

• LoadImage → image input on LMStudioGenerateText
• LMStudioConnection → connection input on LMStudioGenerateText
• LMStudioGenerateText.generated_text → both SaveText|pysssss (writes captions to disk, e.g. 049.txt) and ShowText|pysssss (preview in-graph) in parallel

Notes & limitations

• Only the first image in a batch tensor is sent per call — if you need one API call per image in a batch (e.g. looping ImageStackerDrop's full batch with per-image .txt output), that needs a small extension to the node; ask if you want that added.
• Every call sends the full image as a base64-encoded PNG in the request body — expect slower throughput than a cloud vision API, especially on a modest local GPU.
• The model you select must be vision-capable for image inputs to work; a text-only model will error or silently ignore the image.