Janus Pro

Janus Pro is an advanced unified multimodal AI model developed by DeepSeek, building upon the original Janus model. It integrates multimodal understanding (image-to-text reasoning) and text-to-image generation within a single architecture. Key improvements include an optimized training strategy, expanded training data, and scaling to larger model sizes (1B and 7B parameters). The model employs a novel decoupled visual encoding system that separates the pathways for understanding and generation while sharing a unified Transformer backbone. This design enhances both tasks, achieving competitive performance. In benchmarks, Janus Pro outperforms DALL-E 3 in text-to-image instruction following, with a GenEval score of 0.80 versus DALL-E 3’s 0.67. The model is open-source under the MIT license, making it suitable for research and commercial use. It provides both a Gradio demo for interactive use and API access for integration.

Key Benefits

• Unified multimodal understanding and generation in one model
• Outperforms DALL-E 3 on GenEval benchmark (0.80 vs 0.67)
• Open-source under permissive MIT license
• Decoupled visual encoding improves efficiency and task-specific performance