deepseek-ai/Janus

Decouples visual encoding into separate pathways for understanding and generation while maintaining a unified transformer backbone, enabling single-model performance across both vision-language comprehension and text-to-image synthesis. Janus-Pro and JanusFlow variants integrate rectified flow for improved generation quality, with recent versions demonstrating enhanced instruction-following and generation stability through optimized training strategies and scaled architectures (up to 7B parameters).

17,708 stars. No commits in the last 6 months.