Advancements in diffusion models for high-resolution image and short form video generation

This paper offers an in-depth review of the most recent advancements in diffusion models, particularly highlighting their transformative role in high-resolution image generation and their emerging applications in short-form video generation. Diffusion models, a class of generative models rooted in probabilistic frameworks, have rapidly gained prominence due to their ability to produce photorealistic and detailed outputs by reversing a noise-infusion process. Their strength lies in their capacity to generate high-quality media that exceeds previous limitations of generative models like GANs, especially in terms of diversity and training stability. The study utilized five scientific databases and a systematic search strategy to identify research articles on PubMed, Google Scholar, Scopus, IEEE, and Science Direct relating to the topic. Furthermore, books, dissertations, master's theses, and conference proceedings were utilized in this study. This study encompassed all publications published until 2024. The review begins by delving into the fundamental principles underlying diffusion models, which revolve around the process of gradually adding and removing noise from an image or video over a series of time steps. This section emphasizes the mathematical foundation of diffusion processes, particularly the forward process of noise addition and the reverse process of denoising, which enables these models to generate media with fine detail. A significant portion of this review is dedicated to the impact of diffusion models on high-resolution image and short-form video generation as well as success metrics for evaluating short-form video generation, curation, and summarization, areas where they have been especially transformative. Conclusively, this paper provides a comprehensive exploration of how diffusion models have reshaped the landscape of media generation. From their foundational principles and technical evolution to their applications in high-resolution media and short-form video, the paper highlights both the profound potential of these models and the ongoing challenges that must be addressed for their responsible and scalable use.