DeepSeek stuns tech industry with new AI image generator that beats OpenAI’s DALL-E 3
Chinese artificial intelligence (AI) company DeepSeek unveiled a new image generator soon after its hit chatbot sent shock waves through the tech industry and stock market.
DeepSeek has made headlines for its semi-open-source AI models that rival OpenAI’s ChatGPT despite being made at a fraction of the cost.
On Monday (Jan. 27), DeepSeek claimed that the latest model of its free Janus image generator, Janus-Pro-7B, beat OpenAI’s DALL-E 3 and Stability AI’s Stable Diffusion in benchmark tests, Reuters reported. However, Artificial Analysis, which compares the performance of different AI models, has yet to independently rank DeepSeek’s Janus-Pro-7B among its competitors.
The image generator announcement came at a significant time for DeepSeek and the AI tech industry at large. First, DeepSeek’s free AI assistant chatbot overtook ChatGPT to become the most downloaded free app in Apple’s U.S. App Store. Later that same day, the company announced it was limiting user registrations because of a large-scale cyberattack, though existing users could continue to log in, CNBC reported.
The stock market also reacted to DeepSeek’s low-cost chatbot stardom on Monday. Leading AI chipmaker Nvidia lost $589 billion in stock market value — the biggest one-day market loss in U.S. history (though the company is still worth trillions of dollars).
Related: Scientists propose making AI suffer to see if it’s sentient
DeepSeek’s AI models have taken the tech industry by storm because they use less computing power than typical algorithms and are therefore cheaper to run. The U.S. restricts the number of the best AI computing chips China can import, so DeepSeek’s team developed smarter, more-energy-efficient algorithms that aren’t as power-hungry as competitors, Live Science previously reported.
Janus-Pro-7B is a free model that can analyze and create new images. As with other image generators, users describe in text what image they want, and the image generator creates it. The model’s improvements come from newer training processes, improved data quality and a larger model size, according to a technical report seen by Reuters.
DeepSeek fed the model 72 million high-quality synthetic images and balanced them with real-world data, which reportedly allows Janus-Pro-7B to create more visually appealing and stable images than competing image generators.
DeepSeek runs “open-weight” models, which means users can look at and modify the algorithms, although they don’t have access to its training data.
Source link
