Uni-X: Mitigating Modality Conflict with a Two-End-Separated Architecture for Unified Multimodal Models

Researchers introduce Uni-X, a novel architecture for unified multimodal AI models that addresses gradient conflicts between vision and text processing. The X-shaped design uses modality-specific processing at input/output layers while sharing middle layers, achieving superior efficiency and matching 7B parameter models with only 3B parameters.

• →Uni-X solves gradient conflicts in multimodal transformers by separating initial and final layers for modality-specific processing.
• →The architecture achieves comparable performance to 7B parameter models while using only 3B parameters, demonstrating significant efficiency gains.
• →Uni-X scored 82 on GenEval for image generation while maintaining strong text and vision understanding capabilities.
• →The model identifies that gradient conflicts are most severe in shallow and deep layers, with middle layers naturally aligning semantically.
• →The research provides open-source code and establishes a new foundation for parameter-efficient multimodal AI development.