DeepSeek releases image generation model 'Janus Pro' under MIT license, boasting performance exceeding that of DALL-E 3

Chinese AI startup DeepSeek has released its own image generation model , Janus Pro . Janus Pro is said to outperform OpenAI's DALL-E 3 image generation AI, and is released under the MIT license .

Viral AI company DeepSeek releases new image model family | TechCrunch
https://techcrunch.com/2025/01/27/viral-ai-company-deepseek-releases-new-image-model-family/

DeepSeek has released the image generation model 'Janus Pro' on its AI development platform Hugging Face. Janus Pro's parameter size is between 1 billion and 7 billion, and models with larger parameter sizes perform better. Janus Pro is distributed under the MIT license and can be used commercially.

deepseek-ai/Janus-Pro-1B · Hugging Face
https://huggingface.co/deepseek-ai/Janus-Pro-1B

DeepSeek describes Janus-Pro as a 'novel autoregressive framework' that can analyze and even create new images. DeepSeek explains that 'Janus-Pro outperforms previous integrated models and matches or exceeds the performance of task-specific models' and that 'its simplicity, high flexibility, and effectiveness make Janus-Pro a strong candidate for the next generation of integrated multimodal models.'

The graph below shows the parameter size and average performance of multimodal AI models. It is clear that Janus-Pro's Janus-Pro-7B, which has the largest parameter size, outperforms competing models with similar parameter sizes.

Below is a graph comparing the performance of the AI benchmarks GenEval and DPG-Bench when generating images from text. Janus-Pro-7B outperforms competing models such as DALL-E 3, PixArt-alpha, Emu3-Gen, and Stable Diffusion XL.

TechCrunch, a technology media company, writes, 'Some of the models compared are older versions, and most of the Janus-Pro models can only analyze small images with a maximum resolution of 384 x 384 pixels. Nevertheless, the performance of the Janus-Pro is impressive considering the compactness of the model.'

DeepSeek is funded by quantitative trading firm High-Flyer Capital Management and has risen to the top of the App Store free app rankings, garnering attention from the general public. DeepSeek's language models are trained using computationally efficient techniques, leading industry analysts and engineers to question whether the U.S. can maintain its lead in the AI race and whether demand for AI chips will be sustained.

Chinese AI development company 'DeepSeek' is rapidly emerging as a hot topic in the technology industry, and has also ranked first in the App Store's free app rankings - GIGAZINE

Janus-Pro is also available on GitHub.

GitHub - deepseek-ai/Janus: Janus-Series: Unified Multimodal Understanding and Generation Models
https://github.com/deepseek-ai/Janus