DeepSeek, which began as a modest Chinese startup, continues to amaze the global tech industry. After successfully challenging the dominance of tech giants like Nvidia, OpenAI, and Microsoft in the AI arena, it has set itself a new and ambitious goal: conquering the field of AI-powered image generation.
In a technical paper published this week, the company revealed that its new flagship model, Janus-Pro-7B, has outperformed leading models from OpenAI (DALL-E 3) and Stability AI (Stable Diffusion) in performance tests for generating images from text prompts. The new model is considered a significant upgrade to the Janus model, which was launched at the end of last year.
The company reported that the new model utilizes 72 million high-quality synthetic images, along with real-world data, to generate stable, highly detailed images that align precisely with textual prompts. Additionally, DeepSeek has improved training processes, data quality, and model size, which has now reached 7 billion parameters.
The Janus-Pro-7B model was announced following the success of DeepSeek’s AI chatbot, based on DeepSeek-V3, which became the top-ranked free app in Apple’s App Store in the U.S.
What sets DeepSeek apart from its competitors? Unlike other advanced models that require human-supervised fine-tuning, R1-Zero was trained using a method called Reinforcement Learning. This means the model learned to improve its performance on its own, without human intervention in the process. It’s like teaching a robot to play chess—then letting it train against itself until it becomes a champion, without any human guidance.
This innovative approach not only provides an advanced technological solution but also offers economic efficiency by eliminating the need for massive investments in complex processes. DeepSeek not only conserves resources but also presents a simpler way to achieve remarkable results—marking a potential shift in the AI race.