Janus is a series of open-source multimodal models from DeepSeek AI that unify image understanding and image generation in a single transformer. Released on GitHub, where the repository has more than 17,700 stars, the Janus family — including the original Janus and the improved Janus-Pro at 1B and 7B scales — showed that one autoregressive model can both caption and analyze images and generate new ones from text. The series was last updated in 2025.
The key idea behind Janus is decoupling visual encoding: understanding and generation want different things from an image encoder, so Janus gives each task its own encoder while sharing a single autoregressive transformer as the processing core. This resolves the conflicts that limited earlier unified models and yields strong results on both sides — multimodal benchmarks such as POPE and MME on the understanding side, and GenEval and DPG-Bench on the generation side. Janus-Pro adds scaled-up training data, an optimized training strategy, and larger model sizes, improving instruction following, generation stability, and text-to-image fidelity. Weights, inference code, and a Gradio demo are all included in the repository.
The Janus models and code are free and open source under the MIT license.
Janus is for researchers and developers exploring unified multimodal models, and for teams that want one open model that can both see and create images. Its compact sizes make it a practical baseline for experiments in any-to-any modeling, multimodal chatbots, and creative tools that combine understanding with generation.