neuraLumaDMX

AI-driven translation of natural language into real-time DMX lighting control.

---

Overview

neuraLumaDMX bridges the gap between human intent and DMX512 lighting systems.

Instead of manually programming DMX values, users can input commands like:

"irish flag"
"calm ocean"
"rave strobe"

These are interpreted by a local LLM pipeline, converted into structured lighting data, validated, and sent to physical fixtures via a USB-DMX interface.

---

Features

• Natural Language -> Lighting Control
• Runs fully locally with local LLM (no internet connection required)
• Supports multiple input languages (depending on the LLM)
  • Tested using unsloth/Qwen3.5-9B-GGUF
    • Gaeilge/Irish = "gorm lonrach" (shiny blue)
    • Spanish - "Haz que las luces sean azules" (make the lights blue)
    • Thai - "ไล่ล่าสีเขียวเร็ว" (fast green chase)
• Validation layer prevents invalid DMX output
• Supports 4 unique RGBW fixture outputs
  • An infinite number of RGBW fixtures can be controlled using 4 unique colour settings
• Real time DMX transmission over USB serial connection
• Multiple lighting modes:
  • Static / Strobe / Chase
• Multiple speeds:
  • Slow / Medium / Fast
• Gradio GUI
  • Command input
  • Scene history
  • Saved Scenes
  • Blackout control

---

Pipeline

---

Tech Stack

• Python
• PySerial
• llama.cpp with GGUF LLM (local LLM inference)
• DMX512
• USB-DMX (Enttec Pro)
• JSON
• Gradio (UI)

---

Running the Project

1. Start LLM Server

Any model exposing an OpenAI-compatible endpoint will work. In this project, llama.cpp with unsloth gguf quant of qwen3.5 9b was used

Example (llama.cpp):

llama-server.exe \
  -hf unsloth/Qwen3.5-9B-GGUF \
  --host 127.0.0.1 \
  --port 8033

2. Install Dependencies

pip install gradio pyserial requests

4. Clone this repository:

git clone https://github.com/NigelByrne1/neuraLumaDMX

5. Clone this repository: (if required)

LLM url can be replaced entirely, the port is kept separate for flexibility during development

# USB DMX / serial interface
interface_port = "COM3"
interface_baudrate = 57600

# llama.cpp (or compatible) OpenAI-style API
llm_port = "8033"
llm_url = "http://127.0.0.1:" + llm_port + "/v1/chat/completions"

# DMX start channel (1-based) for each of 4 RGBW fixtures — set fixtures to 4-channel RGBW mode
fixture_start_channels = [1, 5, 9, 13]

6. Run UI (from within repo)

python app.py

Example Commands

"irish flag"
"sunset wash"
"american police lights"
"calm ocean"
"white strobe fast"

---

How It Works

1. Colour Interpretation

First LLM translates user intent into 4 colour names:

"irish flag" becomes "green, white, orange, green"

2. RGBW Conversion

Second LLM converts colours into structured DMX values:

[
  {"r":0,   "g":255, "b":0,   "w":0},   
  {"r":0,   "g":0,   "b":0,   "w":255}, 
  {"r":255, "g":165, "b":0,   "w":0},   
  {"r":0,   "g":255, "b":0,   "w":0}   
]

Validated and parsed before use.

3. Mode & Speed Selection

Third LLM determines behaviour: "rave" becomes "chase, fast"

---

Limitations

• Single DMX universe (512 channels)
• Designed for RGBW fixtures only
• Depends on LLM output quality and consistency

Future Improvements

• Speech to Text
• Add more modes
  • Fade
  • Flicker
  • Pulse
• Explore beyond RGBW fixtures
  • Moving lights
  • Lights with macro settings
• Reduce hardware requirements
  • Smaller LLM or Cloud mode (will increase reliance on internet connection)
  • Increase number of pipelines and code complexity in order to decrease model size reliance