Sitemap

Sander Dieleman made a good case that diffusion is spectral autoregression: noise drowns the high frequencies of an image first, so denoising recovers an image coarse-to-fine, like autoregression over frequency bands. That’s the signal-processing view.

Here I note down some experiments I tried with diffusion models in the past or that I plan on conducting (if i get the compute for it)

I gave the Japanese N1 exam on December 1st 2024, got the result today, passed with a score of 108/180.

A few words I made up. Use them.

I present to the world, the simplest diffusion model! Using only a fully connected feedforward network, and the dead simple sampling function of

My attempt to intuitively explain transformers in brief.