Seeing that I haven't made a post in a while, I thought it'd be nice to make a small update, almost like a devlog. I'm currently working on building music generation models, mostly using diffusion. I should be able to ship a first version of the final tool next month, but in the meantime I trained this tiny model on the MNIST dataset. It's certainly not perfect, but it is small enough to run on CPU, meaning essentailly free to host (but also pretty slow).

If you aren't familiar with diffusion models, it's a type of ML architecture that learns to slowly remove noise from images, thereby generating novel ones. You can look at each step in process with the slider.

Keep a look out for my next post, it should be a cool one!