British AI company Stability AI has announced version 3 of its image generator Stable Diffusion in a Medium variant. Stable Diffusion 3 will be available in different sizes, trained with 800 million to 8 billion parameters. The announced Medium variant was trained with 2 billion parameters. Stability AI provides an API for each, but also makes the source code available for download on Hugging Face’s AI platform.
Advertisement
Stable Diffusion 3 Medium is currently in an early preview state. At the moment, this model is not available to the general public. Interested individuals can be put on a waiting list. Stability AI has the final version of previous models under license Creative ML OpenRails-M Published.
Image Quality and Quick Conversion
The successor to the previous generation should be able to create photorealistic images with a higher level of detail and quality than before. In addition, Stability AI promises that Stable Diffusion 3 should be able to generate text in the image, a promise that several providers made before without being able to deliver reliable results.
In particular, Stable Diffusion 3 should be able to implement “multi-topic signals”, i.e. text entries with several motifs that are related to each other, better than before. This is the weakness of the initial image generators: the precise implementation of concrete inputs such as “a rice bowl with chicken, onions and peas, but without carrots” or “a robot on a hospital bed and a doctor standing next to it in a white coat with a clipboard in his hand” is too much for generative AI to handle from the start.
Stable Diffusion 3 should be able to process complex text inputs more accurately than previous models.
(Image: Sustainability AI)
Criticism and safeguards
In the past, Stability AI in particular has been repeatedly criticized for allowing the training and demonstration of motifs with copyrighted images compared to other providers, which for example block Del-e from OpenAI, Adobe Firefly and MidJourney, including portraits of presidents or the Pope.
Stability AI claims in the announcement of the Stable Diffusion 3: “We have introduced a number of safety measures”, but without delving into what that might mean, the website reports that the manufacturer encourages responsible use.
According to the manufacturer, the new version should also be able to integrate letters into images.
(Image: Sustainability AI)
Licenses and subscriptions
The AI ​​company offers three different subscription models. The non-commercial license is free for individual developers and research. For users with sales under $1 million, Stable Diffusion costs $20 per month. For larger companies, Stability AI offers individual prices.
(AKR)
