Katherine Crowson

Maybe there could be a model wrapper class that has all of the methods that the samplers etc. expect, and the default implementation of these methods just forwards to the...

I need to think about which methods to make standard on the wrapper... * `forward()`/`__call__()` obviously * `sigma_to_t()` and `t_to_sigma()` (for k-diffusion native models these can just be the identity...

I was thinking about something along the lines of the following: ```python class BaseModelWrapper(nn.Module): """The base wrapper class for the k-diffusion model wrapper idiom. Model wrappers should subclass this class...

> hey, sorry for slow reply. > > okay, so this wrapper creates the illusion that the wrapped instance is a subtype of inner_model. seems reasonable (we want the wrapped...

k-diffusion's Euler and Euler Ancestral samplers are just the VE versions of DDIM and the original DDPM sampling method though, they should actually produce the same outputs as their VP...

A callable sounds like a good idea. :)

I didn't want to use a torch.Generator because then you can't do things like "expand"/add details to, or downsample, paths of Brownian motion to try and get a similar path...

I added the callable version in a branch: https://github.com/crowsonkb/k-diffusion/tree/noise-samplers, the ancestral samplers take a `noise_sampler=` argument now that is a callable with two arguments, `sigma` and `sigma_next`, the interval to...

`BrownianTreeNoiseSampler` also supports batches of seeds, `K.sampling.BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=[1, 2, 3, 4])`, seeds must be nonnegative integers.

"New York City, oil on canvas", guidance scale 5.5, all samples drawn with eta=1: 12 steps, DPM++ 2S:  24 steps, DPM++ 2S:  50 steps, Euler ancestral: