This January 2025 paper introduces Janus-Pro, an enhanced artificial intelligence model for multimodal understanding and generation. It builds upon its predecessor, Janus, through optimized training strategies, expanded data, and increased model size. The authors demonstrate that Janus-Pro achieves significant improvements in both multimodal understanding benchmarks and text-to-image generation capabilities, producing more stable and aesthetically pleasing outputs. This work highlights the benefits of decoupling visual encoding for understanding and generation tasks within a unified autoregressive transformer architecture.

Source:

https://arxiv.org/pdf/2501.17811