DeepSeek says its Janus Pro AI model beats rivals in image generation

Audio By Carbonatix

Enlarge

(Reuters) – DeepSeek’s new open-source AI model surpassed Stability AI and Microsoft-backed OpenAI’s models in benchmarks for image generation, the Chinese startup said in a technical report on Monday.

The company said its Janus-Pro-7B AI model outperformed OpenAI’s DALL-E 3 and Stability AI’s Stable Diffusion in a leaderboard ranking for image generation using text prompts.

The new model is an upgrade over Janus, which was launched late last year and comes on the heels of DeepSeek launching a new assistant based on the DeepSeek-V3 model that has become the top-rated free application on Apple’s App Store in the United States.

Tech stocks such as Nvidia and Oracle plummeted on Monday after the China-based company said its DeepSeek-V3 model topped the leaderboard for open-source models.

DeepSeek’s technical report said the new model improves upon Janus by upgrading its training processes, data quality, and model size, resulting in better image stability and richer details.

Janus-Pro achieved more visually appealing and stable image outputs by adding 72 million high-quality synthetic images and balancing them with real-world data, the report added.

The startup added that its larger model version, with up to 7 billion parameters, improved training speed and accuracy in text-to-image generation and task comprehension.

OpenAI and Stability AI did not immediately respond to requests for comment.

(Reporting by Akash Sriram in Bengaluru; Editing by Tasim Zahid and Maju Samuel)