✦
Summary
DeepSeek Janus Pro is a cutting-edge open-weight multimodal AI model developed by DeepSeek, released on January 27, 2025. It unifies multimodal understanding and generation within a single autoregressive transformer, handling both image understanding — looking at a picture and answering questions about it — and text-to-image generation.
Key Features
Unified multimodal understanding and generation in one transformer
Decoupled visual encoding with SigLIP-L for understanding and VQ tokenizer for generation
Superior text-to-image performance, 80% on GenEval
Image analysis identifying objects, relationships, and details
Support for multiple languages with a 4,096-token context window
Open-source availability under the MIT License
1B and 7B model sizes for different hardware
Synthetic aesthetic data for enhanced image generation
Support for fine-tuning on custom datasets
Complex document parsing and video analysis capabilities
About DeepSeek Janus Pro
Janus Pro's key innovation is decoupled visual encoding: it uses SigLIP-L as the vision encoder for understanding and a separate VQ tokenizer for generation, addressing the stability and performance issues of previous multimodal models. It ships in 1B and 7B sizes based on DeepSeek-LLM, with a 4,096-token context window, and incorporates synthetic aesthetic data to enhance text-to-image generation.
On benchmarks, Janus Pro 7B achieved 80% overall accuracy on GenEval, compared to 67% for DALL-E 3 and 74% for Stable Diffusion 3 Medium, and 84.19% on DPG-Bench. It is freely available on GitHub and Hugging Face under the MIT License, with a Gradio demo, making it a compelling option for researchers, developers, and businesses building vision-language applications.