In the future, every microcontroller will be able to run models with 1.5 million parameters.

The Pico-Faces project is an image generation model that can run on a microcontroller. The model implements a latent flow diffusion transformer (DiT). Similar to techniques used by some larger GenAI image models, but with some tricks and efficiencies that pack it down to 5000 times fewer parameters than typical larger image gen models. The model runs on the RP2350 and generates 128x128 pixel images of faces.

The original project supports two output modes: 1) USB serial link with image data transferred in it and rendered on a host PC with help from the Python Imaging Library. 2) VGA output using the Pimoroni Pico VGA Demo Base. I have added support for the Fruit Jam and HSTX/DVI video output.

In the ethos of "Art is what you can get away with", I also added a fun 2x2 color grid effect inspired by Andy Warhol's artwork. 

Parts