A custom node pack that integrates LM Studio's Python SDK with ComfyUI for local LLM/VLM workflows. It provides nodes for model configuration, flexible per-request options, text+vision querying, image captioning, and chat with persistent or temporary conversations.

Requirements

Install the Python dependencies used by these nodes with the requirements.txt or simply install required modules:

pip install lmstudio numpy Pillow

Installation

Open manager and search Manager for WAS LMStudio Easy-Query, or:

Place this folder in ComfyUI/custom_nodes/ComfyUI_LMStudio_EasyQuery.
Start or restart ComfyUI. On startup, this pack clears:
- temp_convos/ (temporary conversations)

Configuration

Edit lmstudio_config.json in the node pack directory to customize default settings:

{
  "temperature": 0.15,
  "max_tokens": 768,
  "seed": 0,
  "unload_after_use": true,
  "default_model": "qwen/qwen2.5-vl-3b",
  "image_max_sizes": [256, 512, 1024, 2048],
  "default_image_max_size": 1024,
  "allowed_root_directories": ["/mnt/k", "/mnt/m"]
}

Configuration Options

| Option | Type | Description | | --- | --- | --- | | temperature | float | Default sampling temperature (0.0-2.0). Lower = more deterministic | | max_tokens | int | Default maximum tokens to generate per response | | seed | int | Default random seed (0 = disabled) | | unload_after_use | bool | Automatically unload model after queries to free VRAM | | default_model | string | Model ID to select by default in the Model node dropdown | | image_max_sizes | array | Available image size options in the Model node dropdown | | default_image_max_size | int | Default maximum edge size for image resizing | | allowed_root_directories | array | Required for dataset nodes. Whitelist of directories that can be accessed |

Adding Allowed Root Directories

For security, the WAS Load Image Directory node requires directories to be explicitly whitelisted. To add directories:

Open lmstudio_config.json
Add your dataset paths to the allowed_root_directories array:

Windows example:

"allowed_root_directories": ["k:/datasets", "c:/data/images", "d:/training"]

Linux/Mac example:

"allowed_root_directories": ["/home/user/datasets", "/mnt/storage/images"]

Notes:

Use forward slashes (/) even on Windows
Subdirectories are automatically allowed (e.g., k:/datasets allows k:/datasets/anime)
Paths are case-sensitive on Linux/Mac
Restart ComfyUI after editing the config

Node Overview

| Node | Type | Key Inputs | Outputs | Notes | | --- | --- | --- | --- | --- | | LM Studio Model | Config | model, manual_model_id, unload_after_use, temperature, max_tokens, seed, image_max_size | LMSTUDIO_MODEL | Produces model config consumed by other nodes. Fetches available models from LM Studio SDK | | LM Studio Options | Options | temperature, max_tokens, seed, top_p, top_k, frequency_penalty, presence_penalty, repeat_penalty, stop | LMSTUDIO_OPTIONS | Per-request overrides; take precedence over Model values | | LM Studio Query | Request | model, mode (one-by-one/batch), system_prompt, user_prompt, images?, options? | responses (list) | Text/VLM single-turn queries. Images resized to image_max_size before encoding | | LM Studio Easy-Caption | Caption | model, images, mode (one-by-one/batch), task_name, user_prompt?, options? | captions (list) | Uses preset tasks from /tasks/*.txt files; same image handling as Query | | LM Studio Chat | Chat | model, conversation_choice, conversation_name, mode, system_prompt, user_prompt, images?, options?, temp_convo | responses (list), queries (list), conversation_name (string) | Persistent conversations stored as JSON; temp_convo stores in temp_convos/ | | WAS Load Image Directory | Dataset | directory_path, recursive, extensions, dataset_output_path, copy_images, force_aspect, max_size, resize_mode | LMSTUDIO_DATASET_IMAGES | Loads images from directory for batch captioning. Requires allowed_root_directories in config | | LM Studio Easy-Caption Dataset | Dataset | model, dataset, task_name, user_prompt?, options? | captions (list), written_caption_paths (list), result (string) | Batch captions entire dataset, writes .txt files alongside images |

Tips

Use a VLM model (e.g., qwen2-vl variants) for image inputs.
Tooltips in the UI describe all fields; Options override Model settings per request.
Enable unload_after_use in the Model node to free VRAM after each query.
Create custom caption tasks by adding .txt files to the /tasks directory.
For dataset captioning, configure allowed_root_directories in lmstudio_config.json for security.

Troubleshooting

Responses look text-only with images:
- Verify a VLM is selected in the Model node
- Ensure images are wired into images input
Conversation selection errors:
- Keep conversation_choice on "New Conversation" and type a name, or pick an existing name from the dropdown
Dataset loading fails:
- Add your dataset directory to allowed_root_directories in lmstudio_config.json
- Example: "allowed_root_directories": ["k:/datasets", "c:/data"]
Model not found:
- Use manual_model_id field if models don't appear in dropdown
- Ensure LM Studio is running and models are downloaded
If the SDK logs websocket shutdown messages on exit, they are informational during normal shutdown.

License

MIT (see repository terms).