Nougan LM Studio 🧠

The local-LLM-in-the-graph pattern won. The KB's essay on LLMs in ComfyUI is blunt about why: an 8B local model isn't writing better English than a frontier API, but it's uncensored, offline, and free per call. NouganLMStudio is that local model, but it doesn't bolt on Ollama or a llama.cpp loader - it talks to LM Studio's dev server, the OpenAI-compatible local server you may already be running. Text in, streamed response out, with vision, audio, and video inputs on the side.

How it works

It calls /v1/chat/completions on your local LM Studio dev server (default http://127.0.0.1:1234), streaming tokens over the ComfyUI websocket to drive an on-node console and progress bar. The interesting part is the media handling: images get downscaled to a longest side (1024/768/512) and base64'd as JPEG content parts, audio becomes base64 WAV via input_audio, and video is either an IMAGE batch of frames or a file path sampled into evenly-spaced frames (video_frames, default 8). You can also drop or paste an image straight onto the node - it uploads to ComfyUI/input/nougan_lms/ and the reference survives save/reload.

There's a ⟳ button that auto-detects loaded models and fills model_name, and a stop button for interrupting. It's an output node, so it can terminate a graph on its own - which is also what saves you from ComfyUI's "prompt has no outputs" validation when you're just chatting.

The inputs that matter

• server_url - LM Studio's dev-server URL. The default is right if you haven't moved it.
• model_name - auto-filled by ⟳, or type it.
• system_prompt / generate_prompt - the system role and your actual message.
• max_tokens - default 512; -1 lets the model decide (the bar goes indeterminate).
• temperature, top_p - the usual sampling knobs.
• stream - token-by-token streaming, on by default; it drives the progress bar.
• image_size, video_frames - how big images get and how many frames video yields.
• prompt_override - a wired STRING replaces generate_prompt. This is how you feed the Prompt Box back in.

Outputs: RESPONSE (the model's answer) and PROMPT (the exact prompt that was sent).

Installing

The suite, one shot - Manager search Nougan, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Winnougan/Nougan_Nodes
# restart ComfyUI

Then in LM Studio: load a model, open the Developer tab, start the server. Sanity check: http://127.0.0.1:1234/v1/models should return JSON. The node itself needs almost nothing extra - one exception is decoding video files, which requires opencv-python; IMAGE batches of frames need nothing. That's the kind of dependency note to read before you blame the node.

Where people get burned

The KB's warning about LLM-in-graph nodes applies here too: any node that reaches the network by design is exactly the shape of thing you should install from a known source - this one is MIT-licensed, open, and read the repo if you care. Practical failures: the model returns the chat habits (role delimiters, preamble) into a raw prompt - that's why the PROMPT output exists, so you can inspect what actually went out. And VRAM: you're now running an LLM and a diffusion model. The node doesn't unload one for the other, so on a tight card budget the quantization of your chat model matters as much as its quality.