Introduction to AI Toolkit

0:03 I literally just pushed this up and I'm building the Docker image to get all of that online right now. So, ironically, this new feature is actually an old feature. In fact, it's actually the very first thing I built AI toolkit to do, and that's these sliders. If you've been following my work for a while, which I doubt anyone's been following me that long, way back in the day during stable fusion 1.5, I came up with a way to train these sliders.

0:26 The basic idea is you can come to some of these and you can adjust the lura weight to change how you know whatever effect you want. On this one it's zoom in. The weights were really high. You had to go like -7 to 7 to get these to work. Things like that. And I released quite a few of these way back in the day. So I did ones for like skin tone, zoom, muscle size, hair length, uh the number of people on and on and on.

0:51 I just kept, you know, making these depth of field, uh, nighttime, curly hair. I built the script. That's what AI toolkit came out. I had this idea to do these and I started doing them. It was originally based on this concept called erasing. And basically, all you do with erasing is you kind of erase a concept from a model. And I was like, okay, if you can erase the concept, then we could enhance the concept. And even better than that, we could invert the Laura and enhance the other way.

1:39 So you end up with a Laura that splits in both directions. And you can run it forward, backward, whatever it is to make these sliders. So, I've been doing this a long time and another paper has come out about this concept after that. In fact, I'm even mentioned in the paper cuz I discussed this with them at the time. So, now I've added all of this to AI toolkit. It's all built into the UI. I'm going to show you how it works.

2:02 I'm going to show you how you can train these today. And the cool thing is this works on every model. You don't really have to build a data set. It's very simple to do. It trains extremely fast. You don't have to teach the model anything new. You're just shifting its knowledge from one concept to another and just making a sliding trajectory along your Laura. So, I'm going to be doing all this locally today. I have two GPUs locally.

2:02 You can tell here I have the GPU 0. This is my 4090. And then I have GPU 1. This is a 5090. A couple cool things about this. We're going to be doing Quinn image today because that's my new favorite model. I absolutely love it. But this will work with anything. It should work with video models, image models, everything. I haven't tested it with video models, but should work should work with stable diffusion 1.5, SDXL, everything pretty much any model that comes out because the way this works is based on how the model itself works with its understanding of the text.

2:50 Now, I'm going to be training on this locally, but if you don't have a good enough GPU at home or maybe you just want some extra compute, you can always pop over to RunPod. I have a video of how to use that um or really any cloud providing service but I do maintain something at runpod to keep that running and if you haven't signed up for runpod yet use my affiliate link in the bottom that just gives me a little extra compute to do some more experimenting on so the big thing that we want to look at normally we are here up here in the top right you see it Laura trainer that's this is a new toggle that exists so when we click on this now we have something called concept slider when you click on that