ComfyUI-LLM-Helper

ComfyUI LLM Helper

A collection of custom nodes for working with LLM APIs in ComfyUI.
It supports everything from API client management to direct ChatCompletions execution and workflow utilities.

Key Features

Dynamic Model Selection

Easily fetch and update the model list directly from your server with a single click of the Update model names button, loaded straight into the combo menu.

<img src="assets/feature_getmodel.webp" />

Stream & Async Execution

Supports real-time streaming for /chat/completions to display generation progress and handle user interrupts instantly. Built on an asynchronous architecture, allowing independent nodes to execute concurrently without blocking your workflow.

<img src="assets/feature_stream_async.webp" />

Dynamic Parameter Control via DynamicCombo

Configure only the parameters you need on the fly without the hassle of chaining multiple node relays.

<img src="assets/feature_llmoptions.webp" />

Installation

cd ComfyUI/custom_nodes
git clone https://github.com/bedovyy/ComfyUI-LLM-Helper
cd ComfyUI-LLM-Helper
pip install -r requirements.txt

Or search and install ComfyUI-LLM-Helper on ComfyUI-Manager.

Usage

Simple ChatCompletions using blueprint

<img src="assets/usage_chatcompletions_simple.webp" />

ChatCompletions with list

<img src="assets/usage_chatcompletions_with_list.webp" />

Nodes

LLM:Client

• Set your LLM server's base_url and select API key from environment variables (loaded securely from ComfyUI/.env).
• Click "Update model names" to query the real /models endpoint and instantly refresh the model_name dropdown with available models, then use the outputs downstream.

LLM:UnloadModel (llama.cpp)

• Sends an unload request to llama-server when running in router mode, freeing the model from memory.

LLM:Unpack client

• Unpacks base_url, api_key, and model_name from the LLM:Client node. This allows these values to be used with other OpenAI API nodes.

Caution: Be aware that the api_key is output as a plain string.

LLM:ChatCompletions

• Sends a request via the OpenAI API /chat/completions endpoint. Requests are streamed by default, allowing you to monitor progress or interrupt the process within the node. (Streaming can be disabled by adding stream: false to LLMCustomJsonOptions.)

LLM:Message

• Constructs messages to be sent to the LLM. In addition to user, system, and assistant roles, custom roles can be selected. When an image is provided as input, an image ID is automatically assigned.

LLM:Options

• Allows specification of parameters such as temperature and top_p.

LLMCustomJsonOptions

• Allows you to include additional request information by writing it in JSON format.

Preview as Text (storable)

• Similar to Preview as Text, but saves the value within the workflow, allowing the stored value to be retrieved and reused.

LLM:ChatCompletions Simple (subgraph blueprint)

• A subgraph blueprint that provides a basic configuration for entering system and user prompts.