latent-diffusion VS wai-conduit

Stable Diffusion XL (SDXL) is the latest text-to-image generation model developed by Stability AI, based on the latent diffusion techniques. SDXL has the potential to create highly realistic images for media, entertainment, education, and industry domains, opening new ways in practical uses of AI imagery.

Here's a link to the early latent-diffusion git, that might be able to create a blank model (I haven't tested it): https://github.com/CompVis/latent-diffusion

Latent diffusion could work for you: https://github.com/CompVis/latent-diffusion (https://arxiv.org/abs/2112.10752)

There are a lot but the one implemented as LDSR in most stable guis is this one. https://github.com/CompVis/latent-diffusion

CompVis/latent-diffusion: High-Resolution Image Synthesis with Latent Diffusion Models (github.com)

The "retrieval based model" refers to https://github.com/CompVis/latent-diffusion#retrieval-augmen..., which uses ScaNN to train a knn embedding searcher.

Stability is basically https://github.com/CompVis/latent-diffusion + training data.

Found relevant code at https://github.com/CompVis/latent-diffusion + all code implementations here

DDIM and PLMS were the original samplers. They were part of Latent Diffusion's repository. They stand for the papers that introduced them, Denoising Diffusion Implicit Models and Pseudo Numerical Methods for Diffusion Models on Manifolds.

yes, https://github.com/CompVis/latent-diffusion

found some context https://github.com/yesodweb/wai/pull/931

Yes. Not because of the developer, but because of how extremely flexible and dynamic the Lisp-family languages are. The power and joy of Lisp is in how it's almost a meta-language, so every project can become its own EDSL. The most famous (infamous?) example of this is Vacietis[2], which is a Common Lisp library that allows C code to be imported directly(!!).

[0] IIRC the Yesod framework's Warp does well on benchmarks, and when you look at code like https://github.com/yesodweb/wai/blob/master/warp/Network/Wai... you can see the lengths they had to go through to work around the choice of implementation language.

[1] Go has a garbage collector, but exposes the stack/heap distinction more directly than Haskell, so it's easier to write allocation-free code in hot paths.

[2] https://github.com/vsedach/Vacietis

My $3/mo vultr box can handle HN loads easily when using a fast backend (I've settled on https://github.com/yesodweb/wai based apps - the only thing that has worked well for me so far).

What kind of metrics do you derive "ton of stuff" from? It seems like the largest blocker is Cryptonite. It's unreasonable to let a handful of packages keep back Nightly. You can now run Warp without it. How does your list of essential blockers for 9.2 look like?

No, Wai is the common abstraction, sort of like connect. Warp is one (only AFAIK) server implementation of it. Both Wai and Warp are developed along side with Yesod, you can find their source code here https://github.com/yesodweb/wai

WAI