Digital Candle: AI-driven flickering flame in 144 LEDs

Digital Candle comes from Kevin Bates, who previously built the credit-card-sized Arduboy handheld and other unusual electronics such as a flexible Tetris game designed to wrap around a magazine. That background places this product in a line of small, playful hardware rather than conventional lighting.

The device's design differs from a typical LED candle in how the flame is produced. Bates trained a small neural network on video of actual candle flames, and the resulting model expresses flame behavior through eight variables that generate the LED lighting patterns. The model uses a convolutional decoder — the same class of component often used for image generation — and runs at 30 frames per second on the ESP32-S3 chip, controlling all 144 LEDs individually.

Because the flame is generated rather than replayed from recorded footage or animation, its motion does not loop on a fixed cycle, and each unit develops its own flicker. The board is also open source, with development software published on GitHub, and because each LED can be controlled separately, it can serve as a clock, a notification display, or even run the Game of Life, a simple cellular automaton. The product avoids an open flame, reducing fire risk, but it requires external power since no rechargeable battery is built in.