<p>Compressive sensing is a signal processing technique that enables the reconstruction of a signal from a small number of measurements, even if the signal is sparse or compressible in a certain domain. Instead of directly acquiring the entire signal, compressive sensing acquires a smaller set of measurements and then reconstructs the original signal using computational algorithms. However, several challenges limit the applicability of existing compressive sensing methods. To overcome the challenges of real-time encoding–decoding in high-frame-rate video compressive sensing a deep learning model with loss function optimization is introduced. Unlike prior methods relying on deep learning and iterative models for reconstruction, this research proposes an Enhanced Generative Adversarial model for video compressive sensing. Initially, the video undergoes conversion into frames and pre-processing involving resizing and normalization to streamline subsequent processing steps. Leveraging these pre-processed frames, compressive video sensing is implemented using the proposed Enhanced Generative Adversarial model. Within the generator model, a novel Residual Dense Spatial Attentive Bidirectional Gated Recurrent Unit is introduced to enhance the performance. Moreover, the Discriminator module is designed by CapsNet architecture. Finally, a loss function optimization strategy is employed through the Chaotic Kookaburra Algorithm to minimize information loss during training, thereby enhancing efficacy. The proposed method outperformed other state-of-art methods in terms of all assessment measures.</p>