DeepSeek launches Janus Pro 7B to rival OpenAI's DALL-E

DeepSeek has made ripples on the internet after rumours surfaced claiming that it outperforms the world's most advanced AI models. Surprisingly, DeepSeek paid less than $6 million to train its AI models, whereas OpenAI committed $100 million. This resulted in DeepSeek becoming the number one free app on the App Store, as well as an unprecedented wipeout of more than $400 billion in NVIDIA's market capitalisation in the United States. The AI chatbot had overwhelming traffic, resulting in DeepSeek server outages and performance issues, and the company has also blamed a cyberattack.

Despite these obstacles, the Chinese AI lab has made significant progress, presenting Janus-Pro, a pioneering open-source AI model. The new model is already generating headlines, as reports show that it outperforms OpenAI's DALL-E, Stability AI's Stable Diffusion, and other picture production models in a variety of benchmarks.

Janus-Pro is an update to Janus, which was released late last year. Janus-Pro is available in a variety of sizes, ranging from the tiny 1 billion parameters to the 7 billion parameter version, which is roughly the size of an SD 3.5L. According to DeepSeek, the largest model, Janus-Pro-7B, performs better than top competitors PixArt-alpha, Emu3-Gen, and SDXL on industry benchmarks GenEval and DPG-Bench. Huggingface, a prominent AI and machine learning website, offers a free download of the Janus-Pro-7B model.

Janus-Pro-7B is built on an autoregressive framework that separates visual encoding processes while maintaining a single transformer architecture for processing. It "not only alleviates the conflict between the visual encoder's roles in understanding and generation but also enhances the framework's flexibility." While Janus-Pro surpasses its competitors across numerous tasks, it does not outperform specialised models designed for specific procedures.

This new image creation model follows DeepSeek's previous success with the R1 language model, which is challenging GPT-4's powers at a fraction of the cost. The low development cost of these advanced models sent shockwaves in the US AI industry.