Open Source Revolution: Stability AI Introduces Stable Diffusion XL 1.0

0:00The wait is over. Dive into Audible's most anticipated collection, The Best of 2025. featuring top audiobooks, podcasts, and originals across all genres. Our editors have carefully curated this year's must-listens from brilliant hidden gems to the buzziest new releases. Every title in this collection has earned its spot. This is your go-to for the absolute best in 2025 audio entertainment. Whether you love thrillers, romance, or nonfiction, your next favorite listen awaits.

0:39A major player has just announced an AI model to compete with MidJourney, and it is open source. So this is going to ruffle a few feathers in the AI space, and I'm thinking make some pretty big impacts on the AI landscape in general. So today on the podcast, we are diving in. So the headline story here is that Stability AI, who is famous for Stable Diffusion, has unveiled Stable Diffusion XL 1.0, which is essentially a really advanced text-to-image model, and this is going to compete directly with Midjourney. So I myself have had the chance to play around with this. You can go find it at clipdrop.co slash Stable Diffusion.

1:19This is actually pretty cool. I was able to generate a hyper-realistic image of a whale on a beach on Vancouver Island. It did a pretty good job. It gave me four additions, variations of this, and this is much more similar to something you would expect out of Mid Journey instead of something that you would expect out of, you know, a lesser model like, let's say, Dolly 2. So overall, I've been fairly impressed with this. Definitely room for improvement in some areas, but this thing seems to have figured out human hands making that work. And the big thing here is that Stable Diffusion is famous for being open source.

1:59What is really cool is that in addition to be able to get this and use this on ClipDrop for, you know, I mean, I used it for free. I played around with it for a minute. You can also get it open source on GitHub. The company's API and consumer apps ClipDrop and Dream Studio also deliver really vibrant, accurate colors and some contrast shadows, lighting, and a lot of other really awesome things that are much better compared to its predecessor. But of course, you know, it is available now open source on GitHub. So in an interview that recently happened with TechCrunch, Stability AI's head of applied machine learnings, that's Joe Penna, described Stable Diffusion XL 1.0 as dynamic, customizable.

2:41It's a model that's ready for fine tuning for concepts and styles. So in addition to being a lot more user friendly, the new model is designed to generate complex images via natural language processing prompts, right? Like you would expect from something like MidJourney. So Stable Diffusion's XL 1.0 encompasses 3.5 billion parameters. It's designed to generate full one megapixel resolution images. It does this in seconds. I've witnessed this, and it does this across multiple aspect ratios. So this is a very considerable improvement over the previous versions, Stable Diffusion XL 0.9, which required a lot more computational power to produce similar high resolution images.

3:22So they've made some really big steps in that regard alone. Stability AI is also kind of touting the fact that they have made a lot of advancements in text generation. So unlike a lot of text to image models that really struggle with generating images with legible text, Stable Diffusion XL 1.0 really excels in advanced text generation and legibility. this is massive for so many reasons if you've ever tried to get a model like um you know dolly 2 to generate you a logo and you're like hey generate me a logo for a company that does x y and z it's gonna like generate this horrible morph thing with these crazy weird like you know letters that aren't actually letters smushed all over it's gonna look terrible that is the problem that they're solving here so i think the model also supports in painting which essentially is you know like reconstructing missing parts of an image and it does outpainting which is extending existing images right so you can throw an image on there and say like continue the image to the right continue the image to the left some people use this to do really cool like panoramic images right because they just keep extending it on both sides till it's essentially a panoramic image um but in any case this is something we've also seen from photoshop um and a number of other i think midjourney's done something here too so you really got to do this to compete nowadays but a really awesome a really awesome move.

4:43And like, let's not forget, this is open source. Absolutely incredible. They also have some image to image prompts, which are really, really cool. So users essentially can input an image, they can add text and a text prompt to that, and then they can create a more detailed variation. So right, you could throw your family photo in there. And then you could say, you know, put the Eiffel Tower in the background or, you know, whatever you want. And it can, it can manipulate the image that you have put in there. In addition to all of these really, really interesting features that they have been able to turn out.

5:16It also understands really complex multi-part instructions provided in short prompts. So this is a pretty big advancement over previous versions that required longer text prompts. Now they're able to cut that down. However, the release of such a powerful open source model doesn't come without a lot of people saying that it's very controversial, right? So stable diffusion XL 1.0, in theory, you know, some people say could be exploited to generate harmful or toxic content. I mean, I'm not sure how that's much different than all the harmful and toxic content that is being generated every day on the internet.

5:55Anyways, you know, maybe someone would argue that that would be accelerated. And, you know, a lot of people are saying the big issue here is that you can create non-consensual deepfakes, right? So I I could say Obama, you know, shooting a gun at a bus or something bad, right? Some famous figure doing something bad. And then I could try to convince people that that was real, right? Creating a deepfake. And the worry here that a lot of people have is you could, you know, theoretically do a lot of that kind of stuff on MidJourney. But MidJourney being a platform that is able to try to, like, kind of crack down.

6:31And if you make a lot of deepfakes on your account, and mostly if they just get famous or go viral, then they'll shut your account down and say, you know, slap on the wrist, bad, don't do that. And so people are worried that if you make a powerful AI model like this, then people are going to, and it's open source, then you can't really control, you can't stop people from generating whatever they want. So I think that's what a lot of people are concerned about or complaining about. So the company actually acknowledges these potential abuses, but they emphasize that it has taken extra steps to mitigate harmful content generation, filtering the model's training data for unsafe imagery and blocking individual problematic terms.

7:12So legal disputes have actually arisen, particularly involving artists and stock photo company Getty Images, and they are essentially protesting against their work being used as training data for generative AI models. So Stability AI claims its activities fall under the fair use doctrine, but it has also pledged to respect opt-out requests from artists and has promised to continue incorporating their requests. So in conjunction with Stable Diffusion XL 1.0's release, Stability AI is launching a fine-tuning feature. It's currently in beta for its API, but this is going to allow users to use as few as five images to specialize generation on specific subject.

7:55So additionally, the company is bringing stability and stable diffusion XL 1.0 to Bedrock, which is Amazon's cloud platform for hosting generative AI models, which is really, you know, they're doing this to further their collaboration with AWS, but this is really massive. Amazon's Bedrock cloud platform is a huge way that companies can integrate AI into what they're building and having stable diffusion XL 1.0 on there, I think is a really, really big step. In addition, I think, I don't want to glaze over the fact that they're doing their fine-tuning, where you can essentially upload images and train it.

8:32If you are a company, and let's say you manufacture air-conditioned parts, and this thing is horrible at generating images of that, you could upload a whole bunch of images. Maybe you've got a database of air-conditioned parts. you upload that whole thing and now all of a sudden it'll be able to generate things uh generate your objects better i think that's really cool um that i think that's a really really big move that not a lot of people are doing and an absolute reason why you'd want something that's open source you want your own model that you can do you know uh some people would say well why won't you just try to get that give that data to someone like um dolly 2 and have them incorporate it obviously that's just a massive process and maybe you want to keep your own data set of images and you have the biggest data set of some sort of subject.

9:21So you would like to be, you know, the one that can, that can, people can generate that on your end. So I think it's really cool. Stability AI's release of Stable Diffusion XL 1.0, I think is a really crucial move for the company. I think that they're experiencing a competitive crunch. So the company actually secured over a hundred million dollars in venture capital, but they've been essentially grappling with cash burn and they're making very aggressive moves to ramp up sales so you know stability ai's ceo imad imad motask remains very optimistic about the company's future he said quote the latest sdxl model represents the next step in stability ai's innovative innovation heritage and ability to bring the most cutting-edge open access models to market for the ai community i think overall all this is really impressive i've tried the tool out i'm i'm quite impressed with it you know the prospect of being able to grab this uh as an open source model and incorporate into projects i'm doing is incredibly appealing incredibly attractive i mean especially when you consider the leading ai generator which is mid journey doesn't even have an api which is absolutely you know it's very annoying for anyone trying to integrate and do image generation so being able to just alone have the API on here I think is a big move and then in addition now they have a lot of these extra features that are just not found anywhere else right fine-tuning with your own images I think that's huge they have a lot of really cool things that they're doing and then making this open source of course is just such a such a massive win right imagine if you had mid-journey open source that would be crazy so I think this is really exciting I'm very excited to follow what they continue to do in the future definitely a pioneer in AI awesome company all around so very excited to watch them grow and see what the next moves are for them