DeepSeek, a Chinese AI startup, has announced that its new open-source model, Janus-Pro-7B, has surpassed leading models from OpenAI (DALL-E3) and Stability AI (Stable Diffusion) in image generation benchmarks.
This achievement follows the success of DeepSeek’s previous model, Janus, and the recent surge in popularity of its DeepSeek-V3-powered assistant on the Apple App Store.
Janus-Pro-7B boasts significant improvements over its predecessor, including enhanced training processes, higher-quality data, and a larger model size. These advancements result in more stable and detailed image outputs. Notably, the inclusion of 72 million high-quality synthetic images, balanced with real-world data, significantly enhances image quality and visual appeal.
The larger 7-billion parameter model also contributes to faster training speeds and improved accuracy in text-to-image generation and task comprehension.
This news has caused a significant stir in the tech world, with stocks of companies like Nvidia and Oracle experiencing a dip following the announcement.
OpenAI and Stability AI have yet to publicly comment on DeepSeek’s claims.
Picture: Deepseek – Github
Overall, I must admit that as a first Open Source Model, it’s good. It may take some time to reach the quality of the current FLUX1 model. My tests showed a bit more detail in FLUX. Since DeepSeek is open-source too, it’s promising and worth following.
Link: https://github.com/deepseek-ai/Janus and https://huggingface.co/deepseek-ai/Janus-Pro-7B
