Libra Architecture Decouples Vision and Language for Multimodal Tasks

The Libra architecture uses switch attention and switch FFN modules to decouple self-modal modeling from cross-modal interaction. This design enables unified performance in both image-to-text understanding and text-to-image generation.