The start-up Stability.ai unveils a third version of its Stable Diffusion image generation model this February 23. “Performance has been significantly improved in multi-topic prompts, image quality and writing capabilities,” specifies the company.
Stable Diffusion 3 uses an architecture of “transformers broadcast”, a technique developed in 2022 and now capable of operating on a large scale. Sora, the tool recently launched by OpenAI and capable of generating videos up to one minute, works on similar principles. This technique relatively new allows better scalability to SB3. The model also uses a “flow matching” technique to improve the quality of what it produces without becoming too resource intensive.
Advertisement
The Stable Diffusion 3 family of models currently includes between 800 million and 8 billion parameters. The objective is to offer users a wide choice to best adapt to their needs. However, they are only accessible in registering on a waiting list For now. This beta phase, as with previous models, should allow Stability.ai to collect initial comments and feedback. More technical details should be given later.
Stability.ai committed to security
The start-up also indicates that it has taken measures to prevent misuse of Stable Diffusion 3 by malicious actors. “Security begins as soon as we train our model and continues throughout testing, evaluation and deployment.” This includes avoiding the creation of “deepfakes” which would contribute to the propagation of false information and images.
Stability.ai is also committed to fighting against any interference taking place during elections due to the misuse of generative artificial intelligence. The agreement, which has 20 signatory companies (including Anthropic, Google, Meta, Microsoft, OpenAI…) is however non-binding and relies entirely on the goodwill of these companies.
The problem of deepfakes is such that the FTC took up the issue last week and announced that it wanted to propose new protection measures to counter false visual representations. The American competition watchdog wants to prohibit any practice that could lead to the theft of the identity of a United States citizen.
Advertisement
Selected for you