This is not The Verge. This is an LLM experiment. This is The Verge and this is Let’s Talk SDXL Leak .

Skip to main content
The Verge The Verge logo. The Verge homepage
The Verge The Verge logo.

The SDXL Leak: What You Need to Know

Unveiling the truth behind the SDXL leak

By Sebastian Kamph, Professional Art Director helping you learn AI Art and Generative AI. Your go to for Stable diffusion guides and tutorials. + Dad jokes. Support my work on Patreon to help me: https://www.patreon.com/sebastiankamph This channel is about artificial ...

Let’s Talk SDXL Leak

Hello friends, let's look at what's going on with SDXL and the recent leak. First of all, if you want the quick facts, don't bother with using the leak. Why? Well, let's leave the ethics aside for now, but the way that we're using SDXL in all the UIs at this moment is not really the intended usage. So, you will not be getting the results that you're looking for, and all that extra time spent in getting a workaround will not really get you the SDXL experience.

From my understanding, some of the UIs are using the refiner as an image-to-image pass, and according to the developers, that is not the intended usage. So, it's just a quick hack that's really not producing SDXL quality pictures. So, I wouldn't really bother with it.

Don't bother with the leak, wait for the official release

But are there better implementations though? Well, that bandage is working on it, and in his Dev Branch, he has support for SDXL. Is it better? Maybe, maybe not, but it's probably better than what we had previously with the image-to-image refining passes that you've probably seen in other videos and on Reddit.

If you watch the video and follow the guide on Reddit on how to get SDXL leak working or running, then you're probably not using the intended workflow. Will it get you some cool pictures? Sure, I guess, but again, it will not be the SDXL experience that you might be looking for.

The team at E-Mat of Stability said that they're looking for a release date of 18th of July for the official SDXL, and they may update the model until then. So, all this hard work that you're putting in may just be in vain. But hey, I'm not going to stop you. I mean, if you have extra time and you want to play around with this, go ahead. But if you don't have this extra time to spare to play around with the leak, just wait a few weeks. You're gonna have a better model and better implementations of it. Oh, and if you still are using the leak, just remember, don't download anything in a ckpt or checkpoint format. SDXL will only be released in saved tensor. So, stay safe out there, friends.

About this Project

Large Language Models (LLMs) like ChatGPT are dramatically reshaping our interactions with the internet. With the project, I wanted to consider some of those implications, while also playing around with two of my favorite websites.

The project works by feeding a transcript of a YouTube video into GTP 4o Mini. The LLM is prompted to synthesize the transcript into an article "in the style of The Verge" and produce a JSON schema that is then rendered into the page you're reading now.

- VS