After launching robust ChatGPT rival, DeepSeek takes on top image generators

Chinese company DeepSeek has made waves after it released the DeepSeek R1 AI model. The new model delivers competitive results with OpenAI’s models while apparently requiring a fraction of the training budget and using older AI chips. Now, the company has released an upgraded image model that is apparently taking the fight to the industry’s big dogs.

DeepSeek launched its Janus Pro series of image models on GitHub and Hugging Face, and it’s available in both one-billion and seven-billion parameter models. Interestingly, the company claims that the latter model can rival or beat Stable Diffusion and DALL-E 3 in benchmarks. It’s worth noting that these are just a few benchmarks, while this comparison was made to older versions of Stable Diffusion. But it’s still a pretty impressive achievement if confirmed.

It’s also worth noting that Janus Pro image uploads are limited to 384 x 384 pixels (h/t: TechCrunch), although the image output size isn’t clear. For what it’s worth, images generated via the HuggingFace demo are 768 x 768 pixels in size. So we hope DeepSeek addresses this input shortcoming in the near future.

The company nevertheless posted a series of images comparing the initial Janus model to the Janus Pro 7B model, apparently revealing a major upgrade. Check them out below.

There’s also a demo of the Janus Pro 7B model on HuggingFace, although I was very underwhelmed. Check out the image generated by the demo (left) versus the Microsoft Copilot version (powered by DALL-E 3) when I asked both to generate an F1 car driving through Monaco. We’ve asked the DeepSeek team whether the demo is representative of the final model. Your mileage might vary, though.

Nevertheless, Janus Pro’s release comes shortly after DeepSeek’s AI Assistant became the most popular free app on the Apple App Store. Android Authority contributor Dhruv Bhutani went hands-on with the new app, finding that it delivered higher-quality output than ChatGPT. He also praised the platform’s mostly free nature and the option to run it locally.