In the realm of video generation, diffusion models have showcased remarkable advancements. However, a lingering challenge persists—the unsatisfactory temporal consistency and unnatural dynamics in inference results. The study explores the intricacies of noise initialization in video diffusion models, uncovering a crucial training-inference gap.  The study addresses challenges in diffusion-based video generation, identifying a training-inference gap…

In the era of edge computing, deploying sophisticated models like Latent Diffusion Models (LDMs) on resource-constrained devices poses a unique set of challenges. These dynamic models, renowned for capturing temporal evolution, demand efficient strategies to navigate the limitations of edge devices. This study addresses the challenge of deploying LDMs on edge devices by proposing a…