Steganography, the practice of concealing information within media, has evolved significantly with advancements in deep learning. This paper presents a novel reversible image steganography framework based on transformer architectures. The proposed method embeds secret messages into the latent representation of an image obtained through a transformer encoder. The decoder, implemented as an inverse transformer network, enables the reconstruction of both the original image and the hidden message. This approach leverages the attention mechanism to enhance feature extraction, allowing for high embedding capacity while maintaining imperceptibility and robustness. Unlike traditional methods, it ensures full reversibility — a critical requirement in domains such as digital forensics and medical imaging. Experimental results demonstrate that the proposed system achieves high peak signal-to-noise ratio (PSNR) and message recovery accuracy, validating its effectiveness and practicality.