Pico-Faces: an image generator that runs on a microcontroller

# What Pico-Faces does

Pico-Faces is an image-generation model adapted to run on very small hardware: the RP2350 microcontroller used in the Raspberry Pi Pico 2. The implementation focuses on generating 128×128 pixel images of human faces using a compact model architecture called a latent flow diffusion transformer (DiT).

# Why this matters for hobbyists and makers

# How it outputs images

Pico-Faces supports two primary output modes:

• USB serial: the microcontroller sends raw image data over USB serial. A host PC receives that data and renders it using Python Imaging Library (PIL) code.
• Video output: the original project supported VGA output using the Pimoroni Pico VGA Demo Base. The author extended support to Fruit Jam and HSTX/DVI video output, allowing the Pico to drive different displays directly.

The author also added a creative 2×2 color-grid effect inspired by Andy Warhol, showing the project can be adapted for artistic display styles as well as technical demonstration.

# Technical approach in plain terms

The model uses a latent diffusion-style approach implemented as a DiT. That means images are represented in a compressed latent space and the diffusion/transformer operations happen there, which reduces memory and compute demands compared with operating directly at full pixel resolution. The write-up emphasizes tricks and efficiencies to shrink the model dramatically while preserving the core image-generation workflow.

# What you can do with it

• Reproduce the demo on RP2350-based hardware (Raspberry Pi Pico 2).
• Send generated images to a host machine over USB for post-processing or archiving.
• Drive displays directly using supported video outputs (VGA, Fruit Jam, HSTX/DVI) to make standalone art or interactive installations.
• Experiment with visual effects, such as the included 2×2 color-grid variant.

# Community and next steps

# Practical constraints to expect

• Output resolution is 128×128, so images are small and best for displays or projects that work with low-resolution artwork.
• Running a compact generative model on RP2350 requires tradeoffs in speed and detail compared to larger models on GPUs.

# Where to learn more

Follow the linked Pico-Faces guide and related Adafruit Learning System resources for implementation details, code examples for serial rendering, and instructions for setting up video output on hardware like Fruit Jam or HSTX/DVI.