Deepsekthe Chinese startup that a few days ago shaken the entire sector AI with its ultra-efficient chatbot, recently launched Janus-proa new family of models Ai able to generate images. According to the data provided by the same company founded by Liang Wenfengthe performance of the most advanced model of the series would exceed those of From 3 Of Openai and other graphics tools based on artificial intelligence, such as Pixart-alpha, Emu3-gen And Stable diffusion xl.
What makes Janus-Pro particularly interesting is his Multimodal architecturewhich allows him not only to create images, but also to analyze them, offering a higher level of versatility than many competing models. Another interesting fact, Janus-Pro was released with mit licensewhich means that it can also be used freely for commercial purposes.
What is and how Janus-pro works: the characteristics
On a technical level, Janus-Pro is available in multiple versions, with Dimensions vary from 1 to 7 billion parameters. The parameters represent the complexity of the model and directly influence the quality of the images generated: more parameters mean, generally, more precise and detailed images.
Deepseek defines Janus-Pro a “Author framework». What does it mean? Basically this indicates that The model can both analyze and create new images.
The tests conducted on the Benchmark Geneval and DPG-Bench show significant results: the Janus-Pro-7b model, the largest of the family, obtains higher scores than some of the main competitors, primarily From 3suggesting excellent efficiency considering its relative compactness. Describing its new model, Deepseek explains:
(Janus-Pro, editor’s note) resolves the limits of previous approaches by removing visual coding in separate paths, while using a single unified transformation architecture for processing. The misunderstanding not only relieves the conflict between the roles of the visual encoder in understanding and generation, but also increases the flexibility of the framework. Janus-Pro exceeds previous unified models and is able to match or overcome the performance of specific models.
As for the model limitsaccording to what was declared by the company, Janus-Pro “Supports the input of images 384 x 384 pixels».
Janus-pro, the AI creates images: how to try it
If you wish Try Janus-Prowe would like to point out that model 7b is freely available for use on the development platform Hugging faceat this link. If you want, it is also possible to download it to make it “turn” locally from this page of Github.