Just to somehow wrap this up:

I did prototype several diffusion models (DDIM, flow matching) with small(ish) u-net and conditional VAEs that could be implemented on a microcontroller, see examples below. But somehow none of the solutions did really "click" where I felt it made sense to go through the pains of implementing this on a MCU. 

One thing is that for these extremely small models, the effective data compressione is not very efficient and very little is gained. Yes, we need scale!

The most realistic option for this use case is simply a trained decoder for image compression and I discussed this already in earlier logs.

As much as I hate loose threads, I will leave this hanging for now. I may pick this up again, once I can think of a way of turning this into an interesting challenge.

Flow matching

Conditional Decoder CNN