What it does
The Janus series has three model families: Janus, JanusFlow and Janus-Pro. Each is built to handle both image understanding and image generation. Janus uses separate visual encoding paths for the two jobs while keeping one transformer. JanusFlow pairs an autoregressive language model with rectified flow for image generation. Janus-Pro, released on 27 January 2025, adds a better training strategy, more training data and larger model sizes. The Hugging Face downloads listed are Janus-1.3B, JanusFlow-1.3B, Janus-Pro-1B and Janus-Pro-7B, each with a 4096 sequence length. The repository has Python examples for multimodal understanding and text-to-image generation, a Gradio demo you can run locally, an online demo on Hugging Face and a FastAPI demo that serves the same functions through an API server. Installation needs Python 3.8 or later. The code repository is under the MIT License. Use of the models is subject to the DeepSeek Model License, which the page says permits commercial usage.
Capabilities
- Image understanding and text-to-image generation in one model family
- Janus-Pro available in 1B and 7B sizes
- Janus, JanusFlow and Janus-Pro weights on Hugging Face
- Gradio and FastAPI demos included
- Code under MIT; models under the DeepSeek Model License
When to use it
- Experimenting with a model that reads images and also generates them
- Projects that fit the DeepSeek Model License, including commercial use
- Hosting a multimodal model behind your own API
