In full hangover for its model R1Deepseek has just launched Janus Pro 7ban AI model to generate images from text and understand other images that are introduced. And yes, it is also open source, although with An asterisk similar to the flame.
Why is it important. Until now, multimodal models have had to juggle between understanding and generation of images, sacrificing efficiency or performance. Janus Pro 7B resolves this dilemma with a new proposal: unifies the understanding and generation of images in a single architecture.
Innovation. The model introduces a “double track” system for visual processing:
- Separate the coding paths to understand and generate images.
- It maintains a single transformer to process all the information.
- Use Siglip-l as visual encoder for 384×384 pixels.
Janus Pro comparative in the face of your predecessor for several applications. Image: Deepseek.
This resolution is its main inconvenience, it seems much more oriented to already experience uses of little ambition than to the applications that we can assume other proposals such as Midjourney either Freepikwhich usually start from 1024×1024 pixels. However, Janus Pro is not a generator of images to use, but a multimodal model with several capacities.
Of course, this resolution allows an optimal balance between quality and processing speed … for uses that are conducted with it.
Between the lines. Janus Pro 7B’s architecture is especially relevant for its efficiency:
- Compact size of 7,000 million (“7b”) of parameters.
- Higher performance to larger specific models.
- Open source under MIT license for the repository, although the model itself requires accepting the Deepseek license.
The MIT license It allows anyone to use, modify and distribute the code freely, even for commercial purposes, provided that the original copyright notice is maintained. It is one of the most permissive licenses that exist.
The Deepseek licenseon the other hand, it is free and allows commercial uses, but includes specific ethical restrictions, such as the prohibition of military use or the generation of misinformation.
In perspective. Janus Pro 7B is not only another multimodal model, but a new paradigm in the architecture of IAS that can see and create. Its unified but decentralized approach may well end up influencing future developments.
The model is built on Deepseek-Llm-7b-Basethe base language model of the Chinese startup, announced in August 2024. of it inherits its language processing capabilities while adding advanced visual abilities. Its 16X subsample system for the generation of images allows you to maintain efficiency without compromising quality.
Outstanding image | Deepseek, Xataka with Mockuuuups Studio
In Xataka | We knew that US Big Tech had a problem with the costs of their AI. Deepseek has just shown to what extent