Show HN: Pico-Faces – A diffusion transformer image generator on a RP Pico 2 MCU (github.com)

A new project has emerged showcasing the capabilities of a generative AI image generator running on the RP2350 microcontroller, commonly found in the Raspberry Pi Pico 2. This innovation allows for the generation of 128×128 RGB images of human faces in a remarkable time frame of 5-20 seconds. The model employs a latent flow diffusion transformer architecture, featuring two variants with only 1.7 million to 2.9 million parameters—significantly fewer than standard diffusion models, which typically require tens of millions. This miniature model can generate images conditionally based on parameters such as gender and facial expressions, showcasing unexpected sophistication. The significance of this development lies in its demonstration that effective image generation can be achieved with minimal computational resources, challenging the norms of high-parameter diffusion models. This project optimizes techniques used in larger models, enabling the smaller architecture to avoid common pitfalls such as collapsing biases during image generation. By successfully scaling with adjustable parameters like steps and classifier-free guidance, it not only revamps expectations for microcontroller capabilities but also opens avenues for low-cost, portable AI solutions in image processing, making advanced AI more accessible to a broader audience.