Development of diffusion models for video sequence generation

Abstract

This paper presents a software implementation for generating short video sequences on a central processing unit. A unified interface is developed for training a compact pixel-space diffusion model and applying pretrained Stable Video Diffusion and CogVideoX models. The educational model uses three-dimensional convolutions, diffusion-timestep conditioning, and noise-prediction error minimization. Stochastic denoising diffusion probabilistic model and deterministic denoising diffusion implicit model sampling are implemented. The mathematical formulation, network architecture, execution parameters, export procedures, and diagnostic evaluation are described. The architecture of the software pipeline for generating video sequences on a CPU is presented. The system includes three alternative branches: training and application of a compact diffusion model, image-to-video generation, and text-to-video generation. All branches share export, evaluation, and computational resource monitoring procedures. The dependence of the parameter count on network width is derived analytically, and the number of denoiser calls is compared across sampling schemes. A protocol is proposed for assessing the trade-off between computational cost and video properties. Results are reported for a smoke test of the compact model comprising two training iterations and the generation of four 8×8-pixel frames. The test confirms that the pipeline executes successfully but does not demonstrate the learning of coherent dynamics or an advantage over other models.This work occupies an engineering and methodological niche: it combines training a small pixel-space Denoising Diffusion Probabilistic Model with pretrained Stable Video Diffusion and CogVideoX models while saving outputs and resource reports. Its contribution is a transparent implementation and a reproducible research protocol; superiority over large models requires separate experimental verification.