Midjourney and Stable Diffusion were some of the first modern AI image generators capable of creating realistic images, but recently OpenAI’s DALL-E model has become more popular (partially because it’s free). Stable Diffusion was notable because it allowed users to generate images locally on their PCs. Now, Stable Diffusion 3 has been released, making it more scalable, and powerful, than ever.
Stability AI has unveiled Stable Diffusion 3 (SD3), its latest image-generating AI model aimed at staying competitive amid recent advancements from rivals like OpenAI and Google. SD3 introduces a new architecture promising enhanced performance across various hardware configurations. Leveraging techniques like the “diffusion transformer” and “flow matching,” SD3 ranges from 800 million to 8 billion parameters—the big improvement here being that with a variable number of parameters, it can be used across a variety of hardware and doesn’t actually force you to have an ultra-powerful PC to generate images. You might still need a powerful machine for getting the best results, however. The 8 billion parameter maximum is actually more than what Stable Diffusion XL has, and that was already pretty resource-intensive.
Notably, the company emphasizes safety measures and puts them front and center in this announcement. As AI becomes more powerful, more effort should be put into making sure that it can’t be used in a harmful way, and Stability AI is well aware of this. While we’d wish the company would be a bit more specific in what it’s doing to make sure the AI won’t do harm, it has stated the following on the matter:
We have taken and continue to take reasonable steps to prevent the misuse of Stable Diffusion 3 by bad actors. Safety starts when we begin training our model and continues throughout the testing, evaluation, and deployment. In preparation for this early preview, we’ve introduced numerous safeguards. By continually collaborating with researchers, experts, and our community, we expect to innovate further with integrity as we approach the model’s public release.
If you want to try it out, you’ll have to wait a bit since it’s not actually available yet, nor we have a timeline—it’s almost as if the company rushed to announce this once it saw impending competition from OpenAI and Google. You can sign yourself up in on a waitlist here, though.
Source: Stability AI, TechCrunch
