Janus-Pro

Janus-Pro takes a distinctive route to unified multimodal AI: instead of forcing one visual encoder to serve two jobs, it decouples visual encoding into separate pathways for understanding and generation while keeping a single transformer backbone, resolving the tension that hampered earlier unified models. Released by DeepSeek in January 2025 in 1B and 7B parameter sizes, it improved substantially on the original Janus in both multimodal understanding and text-to-image quality and became one of the most widely downloaded unified generation models on Hugging Face. The repository provides inference code for both tasks plus a technical report, and installation is a pip install from source with a CUDA-capable GPU expected. Licensing is split: the code carries MIT, while the weights fall under the DeepSeek Model License, which permits commercial use subject to use restrictions. With roughly 17,800 GitHub stars, Janus-Pro remains a reference point for research on autoregressive unified understanding and generation, alongside the earlier Janus and JanusFlow checkpoints in the same repository.