<p>The security of digital communications relies on the generation of robust and random keys, a key element in cryptographic systems. Chaotic systems, though appreciated for their unpredictability, have limitations due to the low dimension of the models used and the slow computation when applied at large scale. This paper proposes a hybrid approach combining chaotic systems (Henon, Lorenz, Rössler), Least Squares Generative Adversarial Networks (LSGAN), Transformer models, and Quantum Key Distribution (QKD) to enhance key generation and the encryption of images and videos. In this approach, sequences from chaotic systems serve as training data for the LSGAN, which learns to generate keys with high entropy and an improved random distribution. These keys are then refined by a Transformer model, which, through its multi-head attention mechanism, captures complex dependencies and strengthens the unpredictability of the generated sequences. Finally, the integration of the quantum BB84 protocol in the key distribution process adds a layer of security resistant to quantum attacks. Experimental results show that this hybrid approach outperforms traditional methods based solely on chaotic systems or classic Generative Adversarial Networks (GANs). It significantly enhances both security and computational efficiency, standing out particularly in image encryption tests, offering better resistance to brute-force and statistical analysis attacks. Furthermore, the proposed approach optimizes encryption performance through a scrambling and diffusion mechanism and can be deployed on platforms such as Raspberry Pi for real-time processing of video streams and images. This research opens new perspectives for the security of modern cryptographic systems, combining advanced techniques in artificial intelligence and quantum cryptography for more secure and efficient key generation and image encryption.</p>