Abstract
Text-to-video (T2V) generation technology holds potential to transformmultiple domains such as education, marketing, entertainment, and assistivetechnologies for individuals with visual or reading comprehension challenges,by creating coherent visual content from natural language prompts. From itsinception, the field has advanced from adversarial models to diffusion-basedmodels, yielding higher-fidelity, temporally consistent outputs. Yet challengespersist, such as alignment, long-range coherence, and computational efficiency.Addressing this evolving landscape, we present a comprehensive survey oftext-to-video generative models, tracing their development from early GANs andVAEs to hybrid Diffusion-Transformer (DiT) architectures, detailing how thesemodels work, what limitations they addressed in their predecessors, and whyshifts toward new architectural paradigms were necessary to overcome challengesin quality, coherence, and control. We provide a systematic account of thedatasets, which the surveyed text-to-video models were trained and evaluatedon, and, to support reproducibility and assess the accessibility of trainingsuch models, we detail their training configurations, including their hardwarespecifications, GPU counts, batch sizes, learning rates, optimizers, epochs,and other key hyperparameters. Further, we outline the evaluation metricscommonly used for evaluating such models and present their performance acrossstandard benchmarks, while also discussing the limitations of these metrics andthe emerging shift toward more holistic, perception-aligned evaluationstrategies. Finally, drawing from our analysis, we outline the current openchallenges and propose a few promising future directions, laying out aperspective for future researchers to explore and build upon in advancing T2Vresearch and applications.