Katherine Crowson

Maybe the most useful value for you to compute is bits/dim? See the top Stack Overflow answer to this question for example, which talks about how to turn a continuous...

I need to look at this but I'm sick right now so it might be a few days

The uniform in log snr schedule has been in k-diffusion for a long time, it is https://github.com/crowsonkb/k-diffusion/blob/master/k_diffusion/sampling.py#L26 :) I intended to implement the first order last step when I first...

At some point I am going to work out what the size parameters have to be divisible by and add it to the model class for each model, because I...

It is for [classifier-free diffusion guidance](https://openreview.net/pdf?id=qw8AKxfYbI). The zero embedding means "no prompt", generate unconditionally. (The model was trained with 20% of the input embeddings replaced at random with the zero...

> Is it really necessary to start from a noisy image? Is it possible to directly operate on a clean image (without any noise) via iterative updates? You can use...

We probably want to keep compatibility for CPU users, can you make the model loading function take the device name (and then only do the .half() if it is on...

Oh :/ I may have to fix that at some point because these models probably *are* fast enough to run on CPU with PLMS sampling.

Fixed by https://github.com/crowsonkb/k-diffusion/commit/7621f11f786ec17119ae9cac8e88971b822a4bbe.

What do Stable Diffusion samples look like with and without this trick? I tried it and it seemed to produce really blurry samples instead at say NFE=5... have you considered...