Janus
Open multimodal models that both understand images and generate them, from DeepSeek.
- Stars
- 17.8k
- Forks
- 2.2k
- Licence
- MIT
- Language
- Python
- Last update
- 1.8 years ago
- Type
- AI model / agent
No commits in over a year — check for forks or alternatives before adopting.
What it is
Janus is a family of AI models from DeepSeek that can both read images (answer questions about them) and create images from text, within one unified model design. The README describes the series as Janus, JanusFlow and Janus-Pro, with pretrained weights published on Hugging Face. The repository provides Python code to load these models and run inference.
- Ask questions about an image by passing a picture and a text prompt to the Janus-Pro model
- Generate images from text prompts using the text-to-image example in the README
- Run the multimodal understanding evaluation code via VLMEvalKit, as mentioned in the news section
- Compare the 1.3B and 7B model sizes listed in the Hugging Face table
Great for
- Researchers studying unified vision-language models that handle understanding and generation together
- Developers who want to experiment with image question answering using Python and Hugging Face tools
- People with a CUDA-capable GPU who want to run open model weights locally
- Teams evaluating open models under the published MIT code license and model agreement
Not ideal for
- Non-developers looking for a ready-to-use app, since the README only shows Python code and a Gradio demo is mentioned without a visible setup path
- Users without a GPU, since the README examples call .cuda() and use bfloat16
- Anyone needing a stable, versioned release, since the repo has no formal release listed
How to get it
- 1
Clone and build from source
You'll need: Python 3.8 or newer · PyTorch and transformers (installed via the repo, details not fully shown) · CUDA-capable GPU (the examples call .cuda())
Install the package from a cloned repo
pip install -e . - 2
Run it
- Clone the repository from GitHub
- Run pip install -e . inside the repo folder
- Load a model such as deepseek-ai/Janus-Pro-7B using the Python example in the Quick Start
- Pass an image and a question to the model to get a text answer
- 3
Get the most out of it
- Use the Hugging Face model table to pick a size that fits your hardware (1B or 7B Janus-Pro, or 1.3B Janus and JanusFlow)
- Try the online demos on Hugging Face Spaces before installing locally
- Use the VLMEvalKit integration to reproduce the understanding benchmark results
- Read the Janus-Pro and JanusFlow technical reports linked in the README for model details
Downloads
Strengths
- Single unified model design handles both image understanding and image generation, per the README
- Multiple model sizes are published on Hugging Face, including 1B and 7B Janus-Pro variants
- Code is MIT licensed, and the README states commercial use is permitted under its model terms
Trade-offs
- Requires Python 3.8+ and, per the examples, a CUDA GPU with bfloat16 support
- Model weights fall under a separate Model Agreement license that users must read
- No formal release is listed and the README is truncated, so some setup details are unclear
How it compares
| Project | Stars | Type | Level | Updated | Best for |
|---|---|---|---|---|---|
| Janus (this) | 17.8k | AI model / agent | advanced | 1.8 years ago | Researchers studying unified vision-language models that handle understanding and generation together |
| anything-llm | 67.8k | App | beginner | 6 days ago | People who want a private ChatGPT-style assistant over their own PDFs and text files |
| unilm | 22.2k | Other | advanced | 2 months ago | Researchers comparing pre-trained model families such as MiniLM, BEiT, or LayoutLM |
Stats synced from GitHub. The assessment above was drafted with AI from the project's README on 2026-10-29 and may miss recent changes — check the README before relying on a command. MIT licence: free to use, modify and distribute, including commercially, as long as you keep the copyright notice.