On January 27, DeepSeek announced that its new open-source AI model had outperformed Stability AI and OpenAI in image generation benchmarks, as disclosed in a technical report. DeepSeek's Janus-Pro-7B model surpassed OpenAI's DALL-E 3 and Stability AI's Stable Diffusion in a leaderboard ranking for generating images from text prompts.
The upgraded Janus-Pro-7B model builds on its predecessor, Janus, which was released late last year. The company recently launched a version based on the DeepSeek-V3 model, which has become the top-rated free application on Apple's US App Store.
Following this announcement, tech stocks like Nvidia and Oracle experienced an increase in value. DeepSeek's technical report highlighted enhancements made to the new model, including improved training processes, data quality, and model size, resulting in enhanced image stability and richer details.
The Janus-Pro-7B model achieved superior image outputs by incorporating 72 million high-quality synthetic images and balancing them with real-world data. The startup also noted that the larger model, with up to 7 billion parameters, contributed to improved training speed and accuracy in text-to-image generation and task comprehension.
Requests for comment from OpenAI and Stability AI remained unanswered at the time of reporting.