DeepSeek's new image model looks like another win for cheaper AI
Jan 28, 2025 - zdnet.com
Chinese AI startup DeepSeek has released Janus-Pro, a multimodal text-to-image AI model, shortly after surpassing ChatGPT as the most downloaded free app on the App Store. Janus-Pro, available under an MIT license, is open source and can be accessed via HuggingFace and GitHub. The model comes in various sizes, from 1B to 7B parameters, with the largest version reportedly outperforming established image generators like Stable Diffusion and Dall-E on benchmarks such as GenEval and DPG-Bench. Despite some limitations in smaller models, Janus-Pro is considered competitive, especially due to DeepSeek's lower training costs compared to US-based AI companies. Nvidia has praised the model as an "excellent AI advancement."
DeepSeek's Janus-Pro builds on its predecessor, Janus, and can both create and analyze images. The company's rapid releases have garnered mixed but generally positive first impressions. There are also reports suggesting that DeepSeek's approach may be more energy-efficient than its US counterparts, potentially disrupting the AI industry and affecting large-scale initiatives like Stargate. As more users test Janus-Pro, its impact on the market and its energy efficiency claims will become clearer.
Key takeaways:
DeepSeek released Janus-Pro, a multimodal text-to-image AI model, which is open source and commercially viable.
Janus-Pro-7B reportedly outperforms established image generators like Stable Diffusion and Dall-E on benchmarks.
Janus-Pro's smaller models are limited to analyzing images of 384 x 384 resolution, but the performance is competitive with lower training costs.
DeepSeek's approach may be more energy efficient than US counterparts, potentially impacting AI industry investment strategies.
