Janus-Pro - A multimodal understanding and generation framework excelling in text-to-image tasks.

## Janus-Pro Model Variants Janus-Pro includes two primary variants: - **Janus-Pro-1B**: Based on DeepSeek-LLM-1.5b-base - **Janus-Pro-7B**: Based on DeepSeek-LLM-7b-base Both support 4096-token sequence lengths and multimodal tasks. ## Benchmark Performance Comparison Janus-Pro surpasses both DALL-E 3 and Stable Diffusion in: - **GenEval** and **DPG-Bench** benchmark tests - Text-to-image generation stability - Multimodal understanding capabilities, such as converting math formulas to LaTeX. ## Visual Encoding Technology Janus-Pro employs **SigLIP-L (ViT-L-16-SigLIP-384)** as its visual encoder, which: - Supports **384×384 resolution** image inputs - Uses decoupled visual encoding to separate understanding/generation roles - Enhances performance through 16× downsampling. ## Model Availability Models are available at: - **Hugging Face**: - [Janus-Pro-1B](https://huggingface.co/deepseek-ai/Janus-Pro-1B) - [Janus-Pro-7B](https://huggingface.co/deepseek-ai/Janus-Pro-7B) - **GitHub**: [Installation guide](https://github.com/deepseek-ai/Janus) with Python ≥3.8 requirements. ## Technical Features Key innovations include: 1. **Decoupled visual encoding** for reduced task conflict 2. **Unified transformer architecture** across model sizes 3. **Specialized tokenizer** from FoundationVision/LlamaGen (16× downsampling) 4. Optimized training strategies and expanded datasets. ## Online Demonstration Options Online demos include: - [Janus-Pro-7B Gradio demo](https://huggingface.co/spaces/deepseek-ai/Janus-Pro-7B) - FastAPI-based testing via `python demo/fastapi_client.py`. ## Licenses for Janus-Pro-7B - **Code**: MIT License ([LICENSE-CODE](https://github.com/deepseek-ai/DeepSeek-LLM/blob/HEAD/LICENSE-CODE)) - **Models**: DeepSeek Model License ([LICENSE-MODEL](https://github.com/deepseek-ai/DeepSeek-LLM/blob/HEAD/LICENSE-MODEL)). ## Research Publications Key references: - Janus-Pro Tech Report: [janus_pro_tech_report.pdf](https://github.com/deepseek-ai/Janus/blob/main/janus_pro_tech_report.pdf) - Janus (base model): [arXiv:2410.13848](https://arxiv.org/abs/2410.13848) - JanusFlow: [arXiv:2411.07975](https://arxiv.org/abs/2411.07975). ### Citation sources: - [Janus-Pro](https://hf-mirror.com/deepseek-ai/Janus-Pro-7B) - Official URL Updated: 2025-04-01