I implemented a very tiny image generation model (latent flow transformer) on a RP2350 microcontroller - it can generate 128x128 images of faces [P]
Heat trend
Collecting trend data
The percentage is based on available heat signal, not comment count or independent people.
Its a 2.4-4 million parameter model, quantized to int8, that can be fully executed on the microcontroller in ~20s with the longest generation. The generated image will then be displayed on a monitor or transferred via usb.
Its a latent flow transformer with 12 layers using AdaLN-Zero for conditioning. CFG is also supported and boosted the image quality a lot. The inference engine streams the weight via DMA from the flash while the previous layer is computed. Relu² activation was used to increase sparsity, which the engine can use to skip calculations.
Took a lot of ablations to get it right and I am quite astonished I got so far with so few parameters. Will post the repo below
https://preview.redd.it/psu567et66mh1.png?width=1167&format=png&auto=webp&s=8b69dfe5caf5a0cde79f03f23fb5843c940bf993