Corresponding Author: Mohammad Shahnawaz Shaikh
The diffusion-based generative models have led to tremendous progress in text-to-video generation. This paper introduces a text-to-video synthesis framework that is trained using a combined noise-prediction, temporal-consistency, and text-alignment objective, and consists of a transformer-based text encoder, a CLIP/CVAE latent-alignment module, a conditional 3D U-Net diffusion model, and an explicit temporal-dynamics module, followed by a super-resolution refinement stage. The three complementary experiments are a quantitative comparison with quantitative baselines (GAN video generation-based, diffusion-based) across three metrics (text-alignment, distributional, structural-fidelity), using standard statistical testing (paired t-tests, Benjamini-Hochberg FDR correction, Cohen's d effect sizes), a human-rated comparative case study with the frontier commercial video generation engines (Runway Gen-4.5, Pika 2.0), and a component ablation study, which isolates the contribution of the temporal-dynamics module, the latent-alignment module, and the super-resolution module. Results indicate statistically significant improvement over conventional GAN and diffusion baselines, and a result competitive with commercial systems even with a significant disparity in the number of resources. Each architectural component's contribution to these benefits is discussed and directions are proposed to close the remaining visual fidelity gap with the larger scale commercial models.