Suhail Doshi: prediction

Transcript context

…And so you would never compete with Civit? You would just kind of let people import? Civit's a site where all these things get kind of hosted by the community, right? And so, yeah, we'll often pull down some of the best things there. I think when we have a significantly better model, we will certainly build something that gets closer to that. Again, I go back to saying just I still think this is very nascent. Things are very underpowered, right? Loras are not easy to train. They're easy for an engineer. It sure would be nicer if I could just pick five or six reference images, right? And they might even be five or six different reference images that are not... They're just very different. They communicate a style, but they're actually like... It's like a mood board, right? And you have to be kind of an engineer almost to train these Loras or go to some site and be technically savvy, at least. It seems like it'd be much better if I could say, I love this style. Here are five images and you tell the model, like, this is what I want. And the model gives you something that's very aligned with what your style is, what you're talking about. And it's a style you couldn't even communicate, right? There's no word. You know, if you have a Tron image, it's not just Tron. It's like Tron plus like four or five different weird things. Even cyberpunk can have its like sub-genre, right? But I just think training Loras and doing that is very heavy. So I hope we can do better than that. We have Sharif from Lexica on the podcast before. Both of you have like a landing page with just a bunch of images where you can like explore things.…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.