Theres a method for speeding up diffusion generation (things like stablediffusion) by several orders of magnitude. It's related to particle physics simulations.

I'm just waiting for the researchers to figure it out for themselves.

it's like watching kids break toys.

there are parallels between non-standard activation functions (selu) which solve the vanishing/exploding gradient problem (in preserving variance and standard deviation across layers) and the renormalization done during markovian diffusion.

imagine a random vector of points N, of some distribution M, which we assume is gaussian at timestep t0.

At timestep t1, there exists a series of candidate functions which transform the values of some subset of our set S.

These functions can each be scored based on how closely they preserve variance and the std.

What we mean to find is some function S, which balances staying as close to variance and the std, while maximizing the transformation to the ideal gaussian.

By doing this we no longer have to retain the markov property.

The candidate functions with the input data become the target loss, the distribution becomes the validation loss.

This is a similar approach in practice, to the minimal extended state space algorithm.

Dude, either write a paper to prove it, or write a demo to demonstrate. Don't just call people children, that's rude

@Hazarth I'm not trained in how to write papers in a way that will be accepted.

@Wisecrack That sure doesn't excuse you! I hold you to higher standards than this!

Anyway, Stable diffusion is FOSS, you can pull the repo and apply the changes yourself and see if it really works the way you expect. Written functional code is often much better proof than any paper you could write.

@Hazarth this platform is for calling out children.

@scor yes, but you better back your words with proof, or you're gonna get called instead. This should apply to life everywhere, not just here :P