Unified Transformer Framework for Integrated Language -Vision Understanding and Content Generation

The integration of visual and linguistic reasoning within a unified computational framework remains a fundamental challenge in multimodal artificial intelligence. This work presents Multi-Modal UNIT (MMU), a transformer-based architecture designed to jointly learn from image as well as text modalities through a single-stream attention mechanism. Unlike traditional dual-encoder or late-fusion approaches, MMU employs lightweight modality adapters that enable fine-grained cross-modal interaction while maintaining architectural efficiency. The model is optimized using a hybrid objective that combines contrastive learning for cross-modal understanding with generative learning for language-conditioned reasoning and caption synthesis. Comprehensive evaluations on standard benchmarks, including COCO, Flickr30k, VQAv2, and NLVR2, demonstrate that MMU achieves 92.4% accuracy and an F1-score of 0.93, while maintaining a compact 210-million-parameter design and an average inference time of 70 milliseconds per sample. The results indicate that MMU provides a scalable and efficient pathway toward general-purpose multimodal intelligence, unifying perception, and generation within a single transformer backbone. The complete implementation of this work is publicly available on Zenodo at https://doi.org/10.5281/zenodo.17499887.